diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index a1db41bd4..70bd34c5d 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -85,6 +85,7 @@ jobs:
           pip install -r backend/requirements.txt
           pip install --no-deps chatterbox-tts
           pip install --no-deps hume-tada
+          pip install --no-deps "voxcpm>=2.0.3,<2.1"
           pip install --no-deps omnivoice
 
       - name: Install MLX dependencies (Apple Silicon only)
@@ -294,6 +295,7 @@ jobs:
           pip install -r backend/requirements.txt
           pip install --no-deps chatterbox-tts
           pip install --no-deps hume-tada
+          pip install --no-deps "voxcpm>=2.0.3,<2.1"
           pip install --no-deps omnivoice
 
       - name: Install PyTorch with CUDA 12.8
@@ -457,6 +459,7 @@ jobs:
           pip install -r backend/requirements.txt
           pip install --no-deps chatterbox-tts
           pip install --no-deps hume-tada
+          pip install --no-deps "voxcpm>=2.0.3,<2.1"
           pip install --no-deps omnivoice
 
       - name: Build ROCm server binary (onedir)
diff --git a/Dockerfile b/Dockerfile
index 4451dcf62..f7af519e1 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -63,6 +63,7 @@ RUN if [ "$PYTORCH_VARIANT" = "rocm" ]; then \
 RUN pip install --no-cache-dir --prefix=/install -r requirements.txt
 RUN pip install --no-cache-dir --prefix=/install --no-deps chatterbox-tts
 RUN pip install --no-cache-dir --prefix=/install --no-deps hume-tada
+RUN pip install --no-cache-dir --prefix=/install --no-deps "voxcpm>=2.0.3,<2.1"
 RUN pip install --no-cache-dir --prefix=/install --no-deps omnivoice
 RUN pip install --no-cache-dir --prefix=/install \
     git+https://github.com/QwenLM/Qwen3-TTS.git
diff --git a/README.md b/README.md
index b4db9a4f6..98e789075 100644
--- a/README.md
+++ b/README.md
@@ -87,16 +87,16 @@ Voicebox is built and maintained with [Capy](https://capy.ai?utm_source=voicebox
 
 ## What is Voicebox?
 
-Voicebox is a **local-first AI voice studio** — a free and open-source alternative to **ElevenLabs** and **WisprFlow** in one app. Clone voices from a few seconds of audio, generate speech in 23 languages across 8 TTS engines, dictate into any text field with a global hotkey, and give any MCP-aware AI agent a voice of your choosing.
+Voicebox is a **local-first AI voice studio** — a free and open-source alternative to **ElevenLabs** and **WisprFlow** in one app. Clone voices from a few seconds of audio, generate speech in 30 languages across 10 TTS engines, dictate into any text field with a global hotkey, and give any MCP-aware AI agent a voice of your choosing.
 
 The two cloud incumbents sit on opposite halves of the voice I/O loop — ElevenLabs on output, WisprFlow on input. Voicebox does both, bridges them with a bundled local LLM for refinement and per-profile personas, and runs the whole thing on your machine.
 
 - **Complete privacy** — models, voice data, and captures never leave your machine
-- **8 TTS engines** — Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro
+- **10 TTS engines** — Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro, OmniVoice (non-commercial), and VoxCPM2
 - **Voice cloning and preset voices** — zero-shot cloning from a reference sample, or 50+ curated preset voices via Kokoro and Qwen CustomVoice
-- **23 languages** — from English to Arabic, Japanese, Hindi, Swahili, and more
+- **30 languages** — from English to Arabic, Japanese, Hindi, Thai, Vietnamese, and more
 - **Post-processing effects** — pitch shift, reverb, delay, chorus, compression, and filters
-- **Expressive speech** — paralinguistic tags like `[laugh]`, `[sigh]`, `[gasp]` via Chatterbox Turbo; natural-language delivery control via Qwen CustomVoice; voices described in plain language via Qwen VoiceDesign
+- **Expressive speech** — paralinguistic tags like `[laugh]`, `[sigh]`, `[gasp]` via Chatterbox Turbo; natural-language delivery control via Qwen CustomVoice and VoxCPM2; voices described in plain language via Qwen VoiceDesign
 - **Unlimited length** — auto-chunking with crossfade for scripts, articles, and chapters
 - **Stories editor** — multi-track timeline for conversations, podcasts, and narratives
 - **Voice input** — global dictation hotkey with push-to-talk and toggle modes, accessibility-verified auto-paste on macOS, in-app mic on every text field, Whisper-based STT
@@ -129,7 +129,7 @@ The two cloud incumbents sit on opposite halves of the voice I/O loop — Eleven
 
 ### Multi-Engine Voice Cloning
 
-Eight TTS engines with different strengths, switchable per-generation:
+Ten TTS engines with different strengths, switchable per-generation:
 
 | Engine                      | Languages | Strengths                                                                                                                                |
 | --------------------------- | --------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
@@ -141,11 +141,13 @@ Eight TTS engines with different strengths, switchable per-generation:
 | **Chatterbox Turbo**        | English   | Fast 350M model with paralinguistic emotion/sound tags                                                                                   |
 | **TADA** (1B / 3B)          | 10        | HumeAI speech-language model — 700s+ coherent audio, text-acoustic dual alignment                                                        |
 | **Kokoro**                  | 8         | 50 curated preset voices, tiny 82M model, fast CPU inference                                                                             |
+| **VoxCPM2**                 | 30        | 2B multilingual model, 48kHz output, high-fidelity reference-audio cloning; CUDA or CPU fallback                                        |
+| **OmniVoice**               | 23 exposed | Multilingual cloning and attribute instructions; CC-BY-NC weights, non-commercial use only                                               |
 
 ### Emotions & Paralinguistic Tags
 
 Only **Chatterbox Turbo** interprets paralinguistic tags like `[laugh]` and
-`[sigh]`. Qwen3-TTS, LuxTTS, Chatterbox Multilingual, and HumeAI TADA read them
+`[sigh]`. Qwen3-TTS, LuxTTS, Chatterbox Multilingual, HumeAI TADA, and VoxCPM2 read them
 literally as text.
 
 With **Chatterbox Turbo** selected, type `/` in the text input to open the tag
@@ -393,7 +395,7 @@ Full API documentation available at `http://127.0.0.1:17493/docs`.
 | Frontend      | React, TypeScript, Tailwind CSS                                                 |
 | State         | Zustand, React Query                                                            |
 | Backend       | FastAPI (Python)                                                                |
-| TTS Engines   | Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox, Chatterbox Turbo, TADA, Kokoro |
+| TTS Engines   | Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox, Chatterbox Turbo, TADA, Kokoro, OmniVoice (non-commercial), VoxCPM2 |
 | STT           | Whisper / Whisper Turbo (PyTorch or MLX)                                        |
 | Local LLM     | Qwen3 (0.6B / 1.7B / 4B), shared runtime with TTS / STT                         |
 | MCP Server    | FastMCP mounted at `/mcp` (Streamable HTTP) + bundled stdio shim binary         |
diff --git a/app/src/components/CapturesTab/CapturesTab.tsx b/app/src/components/CapturesTab/CapturesTab.tsx
index 401b8a3c7..81f55abaf 100644
--- a/app/src/components/CapturesTab/CapturesTab.tsx
+++ b/app/src/components/CapturesTab/CapturesTab.tsx
@@ -269,7 +269,7 @@ export function CapturesTab() {
       // override fall through to whatever the backend picks.
       const engine = voice.default_engine as
         | 'qwen' | 'qwen_custom_voice' | 'qwen_voice_design' | 'luxtts' | 'chatterbox'
-        | 'chatterbox_turbo' | 'tada' | 'kokoro' | 'omnivoice'
+        | 'chatterbox_turbo' | 'tada' | 'kokoro' | 'omnivoice' | 'voxcpm2'
         | undefined;
       return apiClient.generateSpeech({
         profile_id: voice.id,
diff --git a/app/src/components/Generation/EngineModelSelector.tsx b/app/src/components/Generation/EngineModelSelector.tsx
index 9d8b04686..db480282c 100644
--- a/app/src/components/Generation/EngineModelSelector.tsx
+++ b/app/src/components/Generation/EngineModelSelector.tsx
@@ -29,6 +29,7 @@ const ENGINE_OPTIONS = [
   { value: 'tada:3B', label: 'TADA 3B Multilingual', engine: 'tada' },
   { value: 'kokoro', label: 'Kokoro 82M', engine: 'kokoro' },
   { value: 'omnivoice', label: 'OmniVoice (non-commercial)', engine: 'omnivoice' },
+  { value: 'voxcpm2', label: 'VoxCPM2', engine: 'voxcpm2' },
 ] as const;
 
 const ENGINE_DESCRIPTIONS: Record<string, string> = {
@@ -41,6 +42,7 @@ const ENGINE_DESCRIPTIONS: Record<string, string> = {
   tada: 'HumeAI, 700s+ coherent audio',
   kokoro: '82M params, CPU realtime, 8 langs',
   omnivoice: '23 languages, voice design. CC-BY-NC weights: non-commercial use only',
+  voxcpm2: '2B params, 48kHz, multilingual cloning',
 };
 
 /** Engines that only support English and should force language to 'en' on select. */
@@ -54,6 +56,7 @@ const CLONING_ENGINES = new Set([
   'chatterbox_turbo',
   'tada',
   'omnivoice',
+  'voxcpm2',
 ]);
 
 /** Engines that synthesise a voice from a written description. */
diff --git a/app/src/components/Generation/FloatingGenerateBox.tsx b/app/src/components/Generation/FloatingGenerateBox.tsx
index b04673c60..e3fc903dc 100644
--- a/app/src/components/Generation/FloatingGenerateBox.tsx
+++ b/app/src/components/Generation/FloatingGenerateBox.tsx
@@ -154,7 +154,8 @@ export function FloatingGenerateBox({
     | 'kokoro'
     | 'qwen_custom_voice'
     | 'qwen_voice_design'
-    | 'omnivoice';
+    | 'omnivoice'
+    | 'voxcpm2';
   useEffect(() => {
     if (selectedProfile?.language) {
       form.setValue('language', selectedProfile.language as LanguageCode);
diff --git a/app/src/components/ServerSettings/ModelManagement.tsx b/app/src/components/ServerSettings/ModelManagement.tsx
index 5f14157bd..4b953460d 100644
--- a/app/src/components/ServerSettings/ModelManagement.tsx
+++ b/app/src/components/ServerSettings/ModelManagement.tsx
@@ -71,6 +71,8 @@ const MODEL_DESCRIPTIONS: Record<string, string> = {
     'Kokoro 82M by hexgrad. Tiny 82M-parameter TTS that runs at CPU realtime. Supports 8 languages with pre-built voice styles. Apache 2.0 licensed.',
   omnivoice:
     'OmniVoice by k2-fsa. Zero-shot voice cloning via a diffusion language model, plus attribute-based voice design (gender, age, pitch, accent). The model covers 600+ languages; Voicebox exposes the 23 in its own language list. 24kHz output. NON-COMMERCIAL: the weights are CC-BY-NC (code is Apache 2.0; the bundled Higgs Audio V2 codec is under the Boson community licence), so do not use output commercially.',
+  voxcpm2:
+    'VoxCPM2 by OpenBMB. A 2B-parameter multilingual TTS model with 48kHz output and high-fidelity reference-audio voice cloning.',
   'qwen-custom-voice-1.7B':
     'Qwen3-TTS CustomVoice 1.7B by Alibaba. 9 premium preset voices with instruct-based style control for tone, emotion, and prosody. Supports 10 languages.',
   'qwen-custom-voice-0.6B':
@@ -420,7 +422,8 @@ export function ModelManagement() {
         m.model_name.startsWith('chatterbox') ||
         m.model_name.startsWith('tada') ||
         m.model_name.startsWith('kokoro') ||
-        m.model_name.startsWith('omnivoice'),
+        m.model_name.startsWith('omnivoice') ||
+        m.model_name.startsWith('voxcpm2'),
     ) ?? [];
   const whisperModels = modelStatus?.models.filter((m) => m.model_name.startsWith('whisper')) ?? [];
   const llmModels = modelStatus?.models.filter((m) => m.model_name.startsWith('qwen3-')) ?? [];
diff --git a/app/src/components/VoiceProfiles/ProfileForm.tsx b/app/src/components/VoiceProfiles/ProfileForm.tsx
index 4099caffa..87808d070 100644
--- a/app/src/components/VoiceProfiles/ProfileForm.tsx
+++ b/app/src/components/VoiceProfiles/ProfileForm.tsx
@@ -76,6 +76,7 @@ const DEFAULT_ENGINE_OPTIONS = [
   { value: 'tada', label: 'TADA' },
   { value: 'kokoro', label: 'Kokoro 82M' },
   { value: 'omnivoice', label: 'OmniVoice (non-commercial)' },
+  { value: 'voxcpm2', label: 'VoxCPM2' },
 ] as const;
 
 function makeProfileSchema(t: (key: string) => string) {
diff --git a/app/src/i18n/locales/en/translation.json b/app/src/i18n/locales/en/translation.json
index 8ab4dec94..7cfbf6b6c 100644
--- a/app/src/i18n/locales/en/translation.json
+++ b/app/src/i18n/locales/en/translation.json
@@ -890,7 +890,7 @@
           "body": "A few seconds of reference audio is enough. Multi-sample support for higher quality when you want it."
         },
         "engines": {
-          "title": "Seven engines, 23 languages.",
+          "title": "Ten engines, 30 languages.",
           "body": "Pick the tradeoff that fits — quality, speed, or multilingual coverage."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/es/translation.json b/app/src/i18n/locales/es/translation.json
index 286590842..876605522 100644
--- a/app/src/i18n/locales/es/translation.json
+++ b/app/src/i18n/locales/es/translation.json
@@ -889,7 +889,7 @@
           "body": "Bastan unos segundos de audio de referencia. Compatible con varias muestras para mayor calidad cuando lo necesites."
         },
         "engines": {
-          "title": "Siete motores, 23 idiomas.",
+          "title": "Diez motores, 30 idiomas.",
           "body": "Elige el equilibrio que encaje: calidad, velocidad o cobertura multilingüe."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/fr/translation.json b/app/src/i18n/locales/fr/translation.json
index 7717d2c85..9d2272e89 100644
--- a/app/src/i18n/locales/fr/translation.json
+++ b/app/src/i18n/locales/fr/translation.json
@@ -889,7 +889,7 @@
           "body": "Quelques secondes d'audio de référence suffisent. Prise en charge multi-échantillons pour une qualité supérieure quand vous le souhaitez."
         },
         "engines": {
-          "title": "Sept moteurs, 23 langues.",
+          "title": "Dix moteurs, 30 langues.",
           "body": "Choisissez le compromis qui vous convient — qualité, vitesse ou couverture multilingue."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/it/translation.json b/app/src/i18n/locales/it/translation.json
index fdd806aea..4355c1388 100644
--- a/app/src/i18n/locales/it/translation.json
+++ b/app/src/i18n/locales/it/translation.json
@@ -889,7 +889,7 @@
           "body": "Bastano pochi secondi di audio di riferimento. Supporto multi-campione per una qualità superiore quando ne hai bisogno."
         },
         "engines": {
-          "title": "Sette motori, 23 lingue.",
+          "title": "Dieci motori, 30 lingue.",
           "body": "Scegli il compromesso più adatto alle tue esigenze tra qualità, velocità o copertura multilingue."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/ja/translation.json b/app/src/i18n/locales/ja/translation.json
index 782a4b307..7501560b6 100644
--- a/app/src/i18n/locales/ja/translation.json
+++ b/app/src/i18n/locales/ja/translation.json
@@ -889,7 +889,7 @@
           "body": "数秒のリファレンス音声があれば十分です。より高い品質を求めるときは複数サンプルにも対応します。"
         },
         "engines": {
-          "title": "7 つのエンジン、23 言語。",
+          "title": "10 のエンジン、30 言語。",
           "body": "品質、速度、多言語対応 — 用途に合ったトレードオフを選べます。"
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/ko/translation.json b/app/src/i18n/locales/ko/translation.json
index 45165e08c..04c168d7c 100644
--- a/app/src/i18n/locales/ko/translation.json
+++ b/app/src/i18n/locales/ko/translation.json
@@ -889,7 +889,7 @@
           "body": "몇 초의 참조 오디오면 충분합니다. 원할 경우 여러 샘플로 더 높은 품질 지원."
         },
         "engines": {
-          "title": "7개 엔진, 23개 언어.",
+          "title": "10개 엔진, 30개 언어.",
           "body": "품질, 속도, 다국어 지원 중 원하는 트레이드오프를 선택하세요."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/pt-BR/translation.json b/app/src/i18n/locales/pt-BR/translation.json
index df53072c7..48a5328d6 100644
--- a/app/src/i18n/locales/pt-BR/translation.json
+++ b/app/src/i18n/locales/pt-BR/translation.json
@@ -889,7 +889,7 @@
           "body": "Alguns segundos de áudio de referência já bastam. Suporte a múltiplas amostras para maior qualidade quando você quiser."
         },
         "engines": {
-          "title": "Sete motores, 23 idiomas.",
+          "title": "Dez motores, 30 idiomas.",
           "body": "Escolha o equilíbrio ideal — qualidade, velocidade ou cobertura multilíngue."
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/zh-CN/translation.json b/app/src/i18n/locales/zh-CN/translation.json
index 83ae9f440..c218c36a4 100644
--- a/app/src/i18n/locales/zh-CN/translation.json
+++ b/app/src/i18n/locales/zh-CN/translation.json
@@ -889,7 +889,7 @@
           "body": "几秒钟的参考音频就足够了。需要更高质量时,支持多样本克隆。"
         },
         "engines": {
-          "title": "七种引擎,23 种语言。",
+          "title": "十种引擎，30 种语言。",
           "body": "选择最合适的取舍——质量、速度,或多语言覆盖。"
         },
         "agentReady": {
diff --git a/app/src/i18n/locales/zh-TW/translation.json b/app/src/i18n/locales/zh-TW/translation.json
index 81986a87f..5566b0b69 100644
--- a/app/src/i18n/locales/zh-TW/translation.json
+++ b/app/src/i18n/locales/zh-TW/translation.json
@@ -889,7 +889,7 @@
           "body": "幾秒鐘的參考音訊就夠了。需要更高品質時也支援多樣本。"
         },
         "engines": {
-          "title": "七種引擎、23 種語言。",
+          "title": "十種引擎、30 種語言。",
           "body": "選擇最符合需求的取捨——品質、速度,或多語言覆蓋。"
         },
         "agentReady": {
diff --git a/app/src/lib/api/types.ts b/app/src/lib/api/types.ts
index 2de56a154..db92485aa 100644
--- a/app/src/lib/api/types.ts
+++ b/app/src/lib/api/types.ts
@@ -80,7 +80,8 @@ export interface GenerationRequest {
     | 'chatterbox_turbo'
     | 'tada'
     | 'kokoro'
-    | 'omnivoice';
+    | 'omnivoice'
+    | 'voxcpm2';
   instruct?: string;
   /** When true and the profile has a personality prompt, input text is rewritten in-character before TTS. */
   personality?: boolean;
diff --git a/app/src/lib/constants/engines.ts b/app/src/lib/constants/engines.ts
index 6cda56b01..bcc9a9669 100644
--- a/app/src/lib/constants/engines.ts
+++ b/app/src/lib/constants/engines.ts
@@ -9,12 +9,14 @@
  * VoiceDesign take free-form prose ("speak in an angry tone"; VoiceDesign
  * layers it on top of the profile's stored design prompt), while OmniVoice
  * takes a closed vocabulary of attributes ("female, low pitch, british
- * accent") and raises on anything outside it.
+ * accent") and raises on anything outside it. VoxCPM2 maps free-form prose
+ * to a parenthesized delivery prefix in the backend.
  */
 export const INSTRUCT_ENGINES: ReadonlySet<string> = new Set([
   'qwen_custom_voice',
   'qwen_voice_design',
   'omnivoice',
+  'voxcpm2',
 ]);
 
 /** Whether the given engine should show and forward the instruct field. */
diff --git a/app/src/lib/constants/languages.ts b/app/src/lib/constants/languages.ts
index c075892ec..a666fcdc1 100644
--- a/app/src/lib/constants/languages.ts
+++ b/app/src/lib/constants/languages.ts
@@ -6,6 +6,7 @@
  * Chatterbox Multilingual supports 23 languages.
  * Chatterbox Turbo is English-only.
  * Kokoro supports 8 languages.
+ * VoxCPM2 supports 30 languages.
  */
 
 /** All languages that any engine supports. */
@@ -20,10 +21,14 @@ export const ALL_LANGUAGES = {
   fr: 'French',
   he: 'Hebrew',
   hi: 'Hindi',
+  id: 'Indonesian',
   it: 'Italian',
   ja: 'Japanese',
+  km: 'Khmer',
   ko: 'Korean',
+  lo: 'Lao',
   ms: 'Malay',
+  my: 'Burmese',
   nl: 'Dutch',
   no: 'Norwegian',
   pl: 'Polish',
@@ -31,7 +36,10 @@ export const ALL_LANGUAGES = {
   ru: 'Russian',
   sv: 'Swedish',
   sw: 'Swahili',
+  th: 'Thai',
+  tl: 'Tagalog',
   tr: 'Turkish',
+  vi: 'Vietnamese',
   zh: 'Chinese',
 } as const;
 
@@ -71,8 +79,7 @@ export const ENGINE_LANGUAGES: Record<string, readonly LanguageCode[]> = {
   kokoro: ['en', 'es', 'fr', 'hi', 'it', 'pt', 'ja', 'zh'],
   qwen_custom_voice: ['zh', 'en', 'ja', 'ko', 'de', 'fr', 'ru', 'pt', 'es', 'it'],
   qwen_voice_design: ['zh', 'en', 'ja', 'ko', 'de', 'fr', 'ru', 'pt', 'es', 'it'],
-  // OmniVoice covers 600+ languages. Listed here are the ones ALL_LANGUAGES
-  // already knows -- widening that list is a separate change.
+  // Keep OmniVoice's validated 23-language subset; VoxCPM2 adds seven more below.
   omnivoice: [
     'ar',
     'da',
@@ -98,6 +105,38 @@ export const ENGINE_LANGUAGES: Record<string, readonly LanguageCode[]> = {
     'tr',
     'zh',
   ],
+  voxcpm2: [
+    'ar',
+    'da',
+    'de',
+    'el',
+    'en',
+    'es',
+    'fi',
+    'fr',
+    'he',
+    'hi',
+    'id',
+    'it',
+    'ja',
+    'km',
+    'ko',
+    'lo',
+    'ms',
+    'my',
+    'nl',
+    'no',
+    'pl',
+    'pt',
+    'ru',
+    'sv',
+    'sw',
+    'th',
+    'tl',
+    'tr',
+    'vi',
+    'zh',
+  ],
 } as const;
 
 /** Helper: get language options for a given engine. */
diff --git a/app/src/lib/hooks/useGenerationForm.ts b/app/src/lib/hooks/useGenerationForm.ts
index 551a2d06a..a1c0f4219 100644
--- a/app/src/lib/hooks/useGenerationForm.ts
+++ b/app/src/lib/hooks/useGenerationForm.ts
@@ -30,6 +30,7 @@ const generationSchema = z.object({
       'tada',
       'kokoro',
       'omnivoice',
+      'voxcpm2',
     ])
     .optional(),
   personality: z.boolean().optional(),
@@ -107,9 +108,11 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
                     ? 'qwen-voice-design-1.7B'
                     : engine === 'omnivoice'
                       ? 'omnivoice'
-                      : engine === 'qwen_custom_voice'
-                        ? `qwen-custom-voice-${data.modelSize}`
-                        : `qwen-tts-${data.modelSize}`;
+                      : engine === 'voxcpm2'
+                        ? 'voxcpm2'
+                        : engine === 'qwen_custom_voice'
+                          ? `qwen-custom-voice-${data.modelSize}`
+                          : `qwen-tts-${data.modelSize}`;
       const displayName =
         engine === 'luxtts'
           ? 'LuxTTS'
@@ -127,13 +130,15 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
                     ? 'Qwen VoiceDesign 1.7B'
                     : engine === 'omnivoice'
                       ? 'OmniVoice'
-                      : engine === 'qwen_custom_voice'
-                        ? data.modelSize === '1.7B'
-                          ? 'Qwen CustomVoice 1.7B'
-                          : 'Qwen CustomVoice 0.6B'
-                        : data.modelSize === '1.7B'
-                          ? 'Qwen TTS 1.7B'
-                          : 'Qwen TTS 0.6B';
+                      : engine === 'voxcpm2'
+                        ? 'VoxCPM2'
+                        : engine === 'qwen_custom_voice'
+                          ? data.modelSize === '1.7B'
+                            ? 'Qwen CustomVoice 1.7B'
+                            : 'Qwen CustomVoice 0.6B'
+                          : data.modelSize === '1.7B'
+                            ? 'Qwen TTS 1.7B'
+                            : 'Qwen TTS 0.6B';
 
       // Check if model needs downloading
       try {
diff --git a/app/src/lib/utils/format.ts b/app/src/lib/utils/format.ts
index 6719c20c5..ed114c16c 100644
--- a/app/src/lib/utils/format.ts
+++ b/app/src/lib/utils/format.ts
@@ -64,6 +64,7 @@ const ENGINE_DISPLAY_NAMES: Record<string, string> = {
   chatterbox_turbo: 'Chatterbox Turbo',
   tada: 'TADA',
   kokoro: 'Kokoro 82M',
+  voxcpm2: 'VoxCPM2',
 };
 
 export function formatEngineName(engine?: string, modelSize?: string): string {
diff --git a/backend/backends/__init__.py b/backend/backends/__init__.py
index fb80ed0e7..7f0c3d5be 100644
--- a/backend/backends/__init__.py
+++ b/backend/backends/__init__.py
@@ -239,6 +239,7 @@ def is_loaded(self) -> bool:
     "tada": "TADA",
     "kokoro": "Kokoro",
     "omnivoice": "OmniVoice",
+    "voxcpm2": "VoxCPM2",
 }
 
 LLM_ENGINES = {
@@ -430,6 +431,19 @@ def _get_non_qwen_tts_configs() -> list[ModelConfig]:
             size_mb=350,
             languages=["en", "es", "fr", "hi", "it", "pt", "ja", "zh"],
         ),
+        ModelConfig(
+            model_name="voxcpm2",
+            display_name="VoxCPM2",
+            engine="voxcpm2",
+            hf_repo_id="openbmb/VoxCPM2",
+            size_mb=4800,
+            supports_instruct=True,
+            languages=[
+                "zh", "en", "ja", "ko", "de", "fr", "ru", "pt", "es", "it",
+                "he", "ar", "da", "el", "fi", "hi", "id", "km", "lo", "ms",
+                "my", "nl", "no", "pl", "sv", "sw", "th", "tl", "tr", "vi",
+            ],
+        ),
         ModelConfig(
             model_name="omnivoice",
             # Non-commercial: the OmniVoice weights are CC-BY-NC (the code is
@@ -441,9 +455,8 @@ def _get_non_qwen_tts_configs() -> list[ModelConfig]:
             hf_repo_id="k2-fsa/OmniVoice",
             size_mb=3300,
             supports_instruct=True,
-            # OmniVoice covers 600+ languages; these are the ones already in
-            # Voicebox's language enum. Widening that enum is a separate change
-            # that touches the API contract and the frontend.
+            # Retain OmniVoice's validated 23-language subset. VoxCPM2 extends
+            # the API's language enum separately.
             languages=[
                 "zh",
                 "en",
@@ -846,6 +859,10 @@ def get_tts_backend_for_engine(engine: str) -> TTSBackend:
             from .qwen_custom_voice_backend import QwenCustomVoiceBackend
 
             backend = QwenCustomVoiceBackend()
+        elif engine == "voxcpm2":
+            from .voxcpm_backend import VoxCPMBackend
+
+            backend = VoxCPMBackend()
         elif engine == "qwen_voice_design":
             from .qwen_voice_design_backend import QwenVoiceDesignBackend
 
diff --git a/backend/backends/voxcpm_backend.py b/backend/backends/voxcpm_backend.py
new file mode 100644
index 000000000..b86aa49b0
--- /dev/null
+++ b/backend/backends/voxcpm_backend.py
@@ -0,0 +1,187 @@
+"""VoxCPM2 TTS backend implementation.
+
+Wraps the native ``voxcpm`` package for local reference-audio cloning.  The
+backend deliberately disables VoxCPM's optional denoiser so Voicebox can keep
+the dependency surface focused on synthesis; reference audio is already
+validated and normalized by the profile pipeline.
+"""
+
+import asyncio
+import inspect
+import logging
+from pathlib import Path
+
+import numpy as np
+
+from .base import (
+    combine_voice_prompts as _combine_voice_prompts,
+    empty_device_cache,
+    get_torch_device,
+    is_model_cached,
+    manual_seed,
+    model_load_progress,
+)
+
+logger = logging.getLogger(__name__)
+
+VOXCPM_HF_REPO = "openbmb/VoxCPM2"
+VOXCPM_MODEL_FILES = ("config.json", "model.safetensors", "audiovae.pth")
+
+
+class VoxCPMBackend:
+    """Native VoxCPM2 backend for cloned voice profiles."""
+
+    def __init__(self):
+        self.model = None
+        self.model_size = "default"
+        self._device: str | None = None
+        self._model_load_lock = asyncio.Lock()
+
+    def _get_device(self) -> str:
+        # Current VoxCPM2 runs on MPS, but a smoke test showed it substantially
+        # slower than CPU on Apple Silicon and upstream has had MPS-specific
+        # AudioVAE failures. Match Voicebox's Chatterbox/TADA policy: use CUDA
+        # where available and a predictable CPU fallback on macOS.
+        return get_torch_device(force_cpu_on_mac=True)
+
+    @property
+    def device(self) -> str:
+        if self._device is None:
+            self._device = self._get_device()
+        return self._device
+
+    def is_loaded(self) -> bool:
+        return self.model is not None
+
+    def _get_model_path(self, model_size: str = "default") -> str:
+        return VOXCPM_HF_REPO
+
+    def _is_model_cached(self, model_size: str = "default") -> bool:
+        return is_model_cached(VOXCPM_HF_REPO, required_files=list(VOXCPM_MODEL_FILES))
+
+    async def load_model(self, model_size: str = "default") -> None:
+        """Download and load VoxCPM2 lazily on the selected device."""
+        if self.model is not None:
+            return
+        async with self._model_load_lock:
+            if self.model is not None:
+                return
+            await asyncio.to_thread(self._load_model_sync)
+
+    async def load_model_async(self, model_size: str = "default") -> None:
+        """Compatibility alias used by a few model-management call sites."""
+        await self.load_model(model_size)
+
+    def _load_model_sync(self) -> None:
+        model_name = "voxcpm2"
+        is_cached = self._is_model_cached()
+
+        with model_load_progress(model_name, is_cached):
+            from voxcpm import VoxCPM
+
+            device = self.device
+            logger.info("Loading VoxCPM2 on %s...", device)
+            self.model = VoxCPM.from_pretrained(
+                VOXCPM_HF_REPO,
+                load_denoiser=False,
+                optimize=device == "cuda",
+                device=device,
+            )
+
+        logger.info("VoxCPM2 loaded successfully on %s", self.device)
+
+    def unload_model(self) -> None:
+        """Unload VoxCPM2 and release accelerator cache."""
+        if self.model is None:
+            return
+        device = self._device
+        del self.model
+        self.model = None
+        self._device = None
+        empty_device_cache(device)
+        logger.info("VoxCPM2 unloaded")
+
+    async def create_voice_prompt(
+        self,
+        audio_path: str,
+        reference_text: str,
+        use_cache: bool = True,
+    ) -> tuple[dict, bool]:
+        """Store reference paths; VoxCPM2 builds its prompt at generation time."""
+        del use_cache
+        return {
+            "ref_audio": str(audio_path),
+            "ref_text": str(reference_text or "").strip(),
+        }, False
+
+    async def combine_voice_prompts(
+        self,
+        audio_paths: list[str],
+        reference_texts: list[str],
+    ) -> tuple[np.ndarray, str]:
+        # Voicebox's profile service persists combined prompts at 24 kHz.  Keep
+        # that convention; VoxCPM2 resamples references to its 16 kHz encoder
+        # rate internally.
+        return await _combine_voice_prompts(audio_paths, reference_texts, sample_rate=24000)
+
+    async def generate(
+        self,
+        text: str,
+        voice_prompt: dict,
+        language: str = "en",
+        seed: int | None = None,
+        instruct: str | None = None,
+    ) -> tuple[np.ndarray, int]:
+        """Generate 48 kHz speech with reference-audio voice cloning."""
+        del language
+        await self.load_model()
+
+        ref_audio = voice_prompt.get("ref_audio")
+        ref_text = str(voice_prompt.get("ref_text") or "").strip()
+        if ref_audio and not Path(ref_audio).exists():
+            raise FileNotFoundError(f"Reference audio not found: {ref_audio}")
+
+        def _generate_sync():
+            if seed is not None:
+                manual_seed(seed, self.device)
+
+            kwargs = {}
+            if ref_audio:
+                # With delivery instructions, use controllable cloning so the
+                # parenthesized style prefix can take effect. Without an
+                # instruction, Voicebox's accurate reference transcript enables
+                # VoxCPM2's highest-fidelity continuation/cloning mode.
+                kwargs["reference_wav_path"] = str(ref_audio)
+                if ref_text and not (instruct and instruct.strip()):
+                    kwargs["prompt_wav_path"] = str(ref_audio)
+                    kwargs["prompt_text"] = ref_text
+
+            generation_text = text
+            if instruct and instruct.strip():
+                generation_text = f"({instruct.strip()}){text}"
+
+            generation_kwargs = {
+                "text": generation_text,
+                "retry_badcase": True,
+                **kwargs,
+            }
+            # PyPI 2.0.3 predates the explicit seed parameter now available on
+            # VoxCPM main. Voicebox still seeds torch above, and forwards the
+            # argument automatically once a newer VoxCPM release exposes it.
+            try:
+                generate_parameters = inspect.signature(
+                    getattr(self.model, "_generate", self.model.generate)
+                ).parameters
+            except (TypeError, ValueError):
+                generate_parameters = {}
+            if seed is not None and "seed" in generate_parameters:
+                generation_kwargs["seed"] = seed
+
+            wav = self.model.generate(**generation_kwargs)
+            audio = np.asarray(wav, dtype=np.float32).squeeze()
+            if audio.ndim != 1 or audio.size == 0:
+                raise ValueError("VoxCPM2 returned empty audio")
+            sample_rate = int(getattr(getattr(self.model, "tts_model", None), "sample_rate", 48000))
+            return audio, sample_rate
+
+        return await asyncio.to_thread(_generate_sync)
diff --git a/backend/build_binary.py b/backend/build_binary.py
index 5e88fd05c..40af116a5 100644
--- a/backend/build_binary.py
+++ b/backend/build_binary.py
@@ -121,6 +121,8 @@ def build_server(cuda=False, rocm=False):
             "--hidden-import",
             "backend.backends.qwen_custom_voice_backend",
             "--hidden-import",
+            "backend.backends.voxcpm_backend",
+            "--hidden-import",
             "backend.backends.qwen_voice_design_backend",
             "--hidden-import",
             "backend.utils.audio",
@@ -188,6 +190,8 @@ def build_server(cuda=False, rocm=False):
             "--hidden-import",
             "qwen_tts",
             "--hidden-import",
+            "voxcpm",
+            "--hidden-import",
             "qwen_tts.inference",
             "--hidden-import",
             "qwen_tts.inference.qwen3_tts_model",
@@ -200,6 +204,8 @@ def build_server(cuda=False, rocm=False):
             "--copy-metadata",
             "qwen-tts",
             "--copy-metadata",
+            "voxcpm",
+            "--copy-metadata",
             "requests",
             "--copy-metadata",
             "transformers",
@@ -217,6 +223,10 @@ def build_server(cuda=False, rocm=False):
             # modeling_qwen3_tts.py — needs physical .py source files bundled
             "--collect-all",
             "qwen_tts",
+            # VoxCPM2's AudioVAE uses @torch.jit.script, which requires the
+            # physical Python source files in frozen builds.
+            "--collect-all",
+            "voxcpm",
             # Fix for pkg_resources and jaraco namespace packages
             "--hidden-import",
             "pkg_resources.extern",
@@ -772,6 +782,8 @@ def build_shim():
         "--exclude-module",
         "kokoro",
         "--exclude-module",
+        "voxcpm",
+        "--exclude-module",
         "misaki",
         "--exclude-module",
         "spacy",
@@ -835,4 +847,3 @@ def build_shim():
         build_shim()
     else:
         build_server(cuda=cli_args.cuda, rocm=cli_args.rocm)
-
diff --git a/backend/models.py b/backend/models.py
index 409a0c964..9d104a9bf 100644
--- a/backend/models.py
+++ b/backend/models.py
@@ -18,7 +18,7 @@ class VoiceProfileCreate(BaseModel):
     name: str = Field(..., min_length=1, max_length=100)
     description: Optional[str] = Field(None, max_length=500)
     language: str = Field(
-        default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|ms|nl|no|pl|sv|sw|tr)$"
+        default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|id|km|lo|ms|my|nl|no|pl|sv|sw|th|tl|tr|vi)$"
     )
     voice_type: Optional[str] = Field(default="cloned", pattern="^(cloned|preset|designed)$")
     preset_engine: Optional[str] = Field(None, max_length=50)
@@ -81,11 +81,11 @@ class GenerationRequest(BaseModel):
 
     profile_id: str
     text: str = Field(..., min_length=1, max_length=50000)
-    language: str = Field(default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|ms|nl|no|pl|sv|sw|tr)$")
+    language: str = Field(default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|id|km|lo|ms|my|nl|no|pl|sv|sw|th|tl|tr|vi)$")
     seed: Optional[int] = Field(None, ge=0)
     model_size: Optional[str] = Field(default="1.7B", pattern="^(1\\.7B|0\\.6B|1B|3B)$")
     instruct: Optional[str] = Field(None, max_length=500)
-    engine: Optional[str] = Field(default=None, pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice)$")
+    engine: Optional[str] = Field(default=None, pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice|voxcpm2)$")
     personality: bool = Field(
         default=False,
         description="When true and the profile has a personality prompt, the input text is rewritten in-character before TTS.",
@@ -352,7 +352,7 @@ class MCPClientBindingResponse(BaseModel):
     profile_id: Optional[str] = None
     default_engine: Optional[str] = Field(
         None,
-        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice)$",
+        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice|voxcpm2)$",
     )
     default_personality: bool = False
     last_seen_at: Optional[datetime] = None
@@ -371,7 +371,7 @@ class MCPClientBindingUpsert(BaseModel):
     profile_id: Optional[str] = None
     default_engine: Optional[str] = Field(
         None,
-        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice)$",
+        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice|voxcpm2)$",
     )
     default_personality: bool = False
 
@@ -390,7 +390,7 @@ class SpeakRequest(BaseModel):
     )
     engine: Optional[str] = Field(
         None,
-        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice)$",
+        pattern="^(qwen|qwen_custom_voice|qwen_voice_design|luxtts|chatterbox|chatterbox_turbo|tada|kokoro|omnivoice|voxcpm2)$",
     )
     personality: Optional[bool] = Field(
         None,
@@ -398,7 +398,7 @@ class SpeakRequest(BaseModel):
     )
     language: Optional[str] = Field(
         None,
-        pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|ms|nl|no|pl|sv|sw|tr)$",
+        pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|id|km|lo|ms|my|nl|no|pl|sv|sw|th|tl|tr|vi)$",
     )
 
 
diff --git a/backend/requirements.txt b/backend/requirements.txt
index f686eb345..6c1c55c24 100644
--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -19,6 +19,13 @@ huggingface_hub>=0.20.0
 # >=0.1.1 for Qwen3TTSModel.generate_voice_design (the qwen_voice_design engine)
 qwen-tts>=0.1.1
 
+# VoxCPM2 runtime dependencies. The voxcpm package itself is installed with
+# --no-deps by setup scripts because its full metadata pulls Gradio, training,
+# ModelScope, and denoising packages that Voicebox does not use.
+einops>=0.8.0
+safetensors>=0.4.0
+tqdm>=4.60.0
+
 # LuxTTS (voice cloning engine)
 # piper-phonemize needs custom index (no PyPI wheels)
 --find-links https://k2-fsa.github.io/icefall/piper_phonemize.html
@@ -43,7 +50,7 @@ pyloudnorm
 # descript-audio-codec is NOT installed — it pulls onnx/tensorboard
 # via descript-audiotools.  A lightweight shim in utils/dac_shim.py
 # provides the only class TADA uses: Snake1d.)
-torchaudio
+torchaudio>=2.5.0
 
 # OmniVoice sub-dependencies (omnivoice itself is installed --no-deps
 # in the setup script because it requires transformers>=5.3, while several
diff --git a/backend/services/history.py b/backend/services/history.py
index 809a556bf..b69d2d47a 100644
--- a/backend/services/history.py
+++ b/backend/services/history.py
@@ -117,7 +117,7 @@ async def create_generation(
         instruct: Natural language instruction used (if any)
         generation_id: Pre-assigned ID (for async generation flow)
         status: Generation status (generating, completed, failed)
-        engine: TTS engine used (qwen, luxtts, chatterbox, chatterbox_turbo)
+        engine: TTS engine used (qwen, luxtts, chatterbox, chatterbox_turbo, voxcpm2)
         model_size: Model size variant (1.7B, 0.6B) — only relevant for qwen
         source: Origin marker stored on the row. ``"manual"`` for regular
             /generate calls; ``"personality_speak"`` for rows created
diff --git a/backend/services/profiles.py b/backend/services/profiles.py
index 22bdf0092..684dd37d1 100644
--- a/backend/services/profiles.py
+++ b/backend/services/profiles.py
@@ -32,7 +32,7 @@
 
 logger = logging.getLogger(__name__)
 
-CLONING_ENGINES = {"qwen", "luxtts", "chatterbox", "chatterbox_turbo", "tada", "omnivoice"}
+CLONING_ENGINES = {"qwen", "luxtts", "chatterbox", "chatterbox_turbo", "tada", "omnivoice", "voxcpm2"}
 
 # Engines that synthesise a voice from a natural-language description
 # (voice_type "designed") rather than from samples or a preset id.
diff --git a/backend/tests/E2E_MODEL_TEST_DESIGN.md b/backend/tests/E2E_MODEL_TEST_DESIGN.md
index 0e6b49fb2..bc6d17f2f 100644
--- a/backend/tests/E2E_MODEL_TEST_DESIGN.md
+++ b/backend/tests/E2E_MODEL_TEST_DESIGN.md
@@ -4,7 +4,7 @@
 
 A single script, runnable on macOS and Windows, that exercises every TTS model against the **frozen PyInstaller binary** (not the dev server), captures per-model pass/fail and error messages, and exits non-zero if any model fails. Generation is strictly sequential — one model loaded at a time.
 
-## Test matrix (10 runs)
+## Test matrix (12 runs)
 
 Derived from `backend/backends/__init__.py:185-316`. Each row maps to one `POST /generate` call.
 
@@ -20,8 +20,10 @@ Derived from `backend/backends/__init__.py:185-316`. Each row maps to one `POST
 | 8 | `tada`                | `1B`       | cloned       | tada-1b, English only |
 | 9 | `tada`                | `3B`       | cloned       | tada-3b-ml, multilingual |
 | 10| `kokoro`              | —          | preset       | `preset_voice_id="af_heart"` |
+| 11| `voxcpm2`             | —          | cloned       | shared reference WAV + transcript |
+| 12| `qwen_voice_design`   | —          | designed     | single 1.7B checkpoint, stored design prompt |
 
-Cloned engines (1, 2, 5, 6, 7, 8, 9) share **one** profile created once with the reference WAV. Preset profiles are created separately, one for kokoro and one for qwen_custom_voice.
+Cloned engines (1, 2, 5, 6, 7, 8, 9, 11) share **one** profile created once with the reference WAV. Preset profiles are created separately, one for kokoro and one for qwen_custom_voice. VoiceDesign uses a separate designed profile. OmniVoice is not yet included in this frozen-binary matrix.
 
 Language for every run: `en` (covers every engine's supported set).
 
diff --git a/backend/tests/fixtures/README.md b/backend/tests/fixtures/README.md
index 341b98acb..32b3b363b 100644
--- a/backend/tests/fixtures/README.md
+++ b/backend/tests/fixtures/README.md
@@ -5,7 +5,7 @@ Place two files here before running `test_all_models_e2e.py`:
 - `reference_voice.wav` — a clean speech sample, mono, 16–24 kHz, ~5–15 seconds.
 - `reference_voice.txt` — the **exact** transcription of the WAV (single line, no trailing newline required).
 
-These are used to create a cloned voice profile for every cloning-capable engine (qwen, luxtts, chatterbox, chatterbox_turbo, tada). Keep them out of version control if they contain personal audio — this directory is not gitignored by default, so add them to `.gitignore` locally if needed.
+These are used to create a cloned voice profile for every cloning-capable engine (qwen, luxtts, chatterbox, chatterbox_turbo, tada, voxcpm2). Keep them out of version control if they contain personal audio — this directory is not gitignored by default, so add them to `.gitignore` locally if needed.
 
 You can point the test at different files with:
 
diff --git a/backend/tests/test_all_models_e2e.py b/backend/tests/test_all_models_e2e.py
index 1d9b69bdf..c91f3d515 100644
--- a/backend/tests/test_all_models_e2e.py
+++ b/backend/tests/test_all_models_e2e.py
@@ -63,6 +63,7 @@ class MatrixRow:
     MatrixRow("tada 1B",                "tada",              "1B",   "cloned",          "tada-1b"),
     MatrixRow("tada 3B",                "tada",              "3B",   "cloned",          "tada-3b-ml"),
     MatrixRow("kokoro",                 "kokoro",            None,   "preset_kokoro",   "kokoro"),
+    MatrixRow("voxcpm2",                "voxcpm2",           None,   "cloned",          "voxcpm2"),
 ]
 
 TEXT = "The quick brown fox jumps over the lazy dog."
diff --git a/backend/tests/test_generation_engine_resolution.py b/backend/tests/test_generation_engine_resolution.py
index 7cd3dc344..74dc0bf84 100644
--- a/backend/tests/test_generation_engine_resolution.py
+++ b/backend/tests/test_generation_engine_resolution.py
@@ -25,3 +25,22 @@ def test_explicit_engine_is_preserved():
 def test_invalid_explicit_engine_is_rejected():
     with pytest.raises(ValidationError):
         _request(engine="invalid")
+
+
+@pytest.mark.parametrize("engine", ["voxcpm2", "omnivoice", "qwen_voice_design"])
+def test_omitted_engine_uses_saved_profile_preference(engine):
+    from types import SimpleNamespace
+
+    from backend.routes.generations import _resolve_generation_engine
+
+    profile = SimpleNamespace(default_engine=engine, voice_type="cloned", preset_engine=None)
+    assert _resolve_generation_engine(_request(), profile) == engine
+
+
+@pytest.mark.parametrize("engine", ["voxcpm2", "omnivoice"])
+def test_cloned_profiles_accept_both_multilingual_engines(engine):
+    from types import SimpleNamespace
+
+    from backend.services.profiles import validate_profile_engine
+
+    validate_profile_engine(SimpleNamespace(voice_type="cloned"), engine)
diff --git a/backend/tests/test_models_load_by_name.py b/backend/tests/test_models_load_by_name.py
index 1e1f50a32..41c224254 100644
--- a/backend/tests/test_models_load_by_name.py
+++ b/backend/tests/test_models_load_by_name.py
@@ -78,7 +78,7 @@ def test_load_by_model_name_targets_that_model(client, loaded):
 
 @pytest.mark.parametrize(
     "model_name",
-    ["qwen-tts-0.6B", "chatterbox-tts", "whisper-turbo", "qwen3-0.6b"],
+    ["qwen-tts-0.6B", "chatterbox-tts", "whisper-turbo", "qwen3-0.6b", "voxcpm2", "omnivoice", "qwen-voice-design-1.7B"],
 )
 def test_load_dispatches_across_engine_families(client, loaded, model_name):
     """TTS, STT, and LLM entries all resolve through the same registry."""
diff --git a/backend/tests/test_openai_compat.py b/backend/tests/test_openai_compat.py
index 847a529c4..e65b49b1c 100644
--- a/backend/tests/test_openai_compat.py
+++ b/backend/tests/test_openai_compat.py
@@ -146,9 +146,10 @@ async def test_speech_wav_default_with_openai_alias_uses_profile_engine(client,
     assert calls["voice_prompt_engine"] == "qwen"
 
 
-async def test_speech_model_is_a_voicebox_engine(client, app):
-    await client.audio.speech.create(model="chatterbox", voice="Morgan", input="Hi")
-    assert app.state.calls["load"] == ("chatterbox", "default")
+@pytest.mark.parametrize("engine", ["chatterbox", "voxcpm2", "omnivoice"])
+async def test_speech_model_is_a_voicebox_engine(client, app, engine):
+    await client.audio.speech.create(model=engine, voice="Morgan", input="Hi")
+    assert app.state.calls["load"] == (engine, "default")
 
 
 async def test_speech_model_variant_selects_size(client, app):
diff --git a/backend/tests/test_voxcpm_backend.py b/backend/tests/test_voxcpm_backend.py
new file mode 100644
index 000000000..d4f1b51bd
--- /dev/null
+++ b/backend/tests/test_voxcpm_backend.py
@@ -0,0 +1,139 @@
+"""Unit tests for the native VoxCPM2 backend without loading model weights."""
+
+import asyncio
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+from backend.backends.voxcpm_backend import VoxCPMBackend
+
+
+@pytest.mark.asyncio
+async def test_create_voice_prompt_preserves_reference_audio_and_text(tmp_path):
+    backend = VoxCPMBackend()
+    audio_path = tmp_path / "reference.wav"
+
+    prompt, cached = await backend.create_voice_prompt(str(audio_path), "  hello  ")
+
+    assert prompt == {"ref_audio": str(audio_path), "ref_text": "hello"}
+    assert cached is False
+
+
+@pytest.mark.asyncio
+async def test_generate_forwards_ultimate_clone_prompt_and_returns_48khz_audio(tmp_path):
+    backend = VoxCPMBackend()
+    audio_path = tmp_path / "reference.wav"
+    audio_path.write_bytes(b"reference")
+    backend._device = "cpu"
+    captured = {}
+
+    class FakeModel:
+        tts_model = SimpleNamespace(sample_rate=48000)
+
+        def _generate(self, text, seed=None, **kwargs):
+            pass
+
+        def generate(self, **kwargs):
+            captured.update(kwargs)
+            return np.zeros(32, dtype=np.float32)
+
+    backend.model = FakeModel()
+
+    audio, sample_rate = await backend.generate(
+        "hello",
+        {"ref_audio": str(audio_path), "ref_text": "reference text"},
+        seed=7,
+        instruct="warm and clear",
+    )
+
+    assert audio.dtype == np.float32
+    assert audio.shape == (32,)
+    assert sample_rate == 48000
+    assert captured == {
+        "text": "(warm and clear)hello",
+        "seed": 7,
+        "retry_badcase": True,
+        "reference_wav_path": str(audio_path),
+    }
+
+
+@pytest.mark.asyncio
+async def test_generate_uses_transcript_guided_cloning_without_instruction(tmp_path):
+    backend = VoxCPMBackend()
+    audio_path = tmp_path / "reference.wav"
+    audio_path.write_bytes(b"reference")
+    captured = {}
+
+    class LegacyModel:
+        tts_model = SimpleNamespace(sample_rate=48000)
+
+        def _generate(self, text, **kwargs):
+            pass
+
+        def generate(self, **kwargs):
+            captured.update(kwargs)
+            return np.ones(16, dtype=np.float32)
+
+    backend.model = LegacyModel()
+    backend._device = "cpu"
+
+    await backend.generate("hello", {"ref_audio": str(audio_path), "ref_text": "reference text"}, seed=9)
+
+    assert captured == {
+        "text": "hello",
+        "retry_badcase": True,
+        "reference_wav_path": str(audio_path),
+        "prompt_wav_path": str(audio_path),
+        "prompt_text": "reference text",
+    }
+
+
+def test_model_registry_contains_voxcpm2():
+    from backend.backends import get_model_config, get_tts_backend_for_engine
+
+    config = get_model_config("voxcpm2")
+    assert config is not None
+    assert config.engine == "voxcpm2"
+    assert config.hf_repo_id == "openbmb/VoxCPM2"
+    assert config.supports_instruct is True
+    assert len(config.languages) == 30
+    assert isinstance(get_tts_backend_for_engine("voxcpm2"), VoxCPMBackend)
+
+
+def test_voxcpm2_request_accepts_languages_added_by_the_engine():
+    from backend.models import GenerationRequest, VoiceProfileCreate
+
+    request = GenerationRequest(profile_id="profile", text="สวัสดี", language="th", engine="voxcpm2")
+    profile = VoiceProfileCreate(name="Thai voice", language="th")
+
+    assert request.language == "th"
+    assert profile.language == "th"
+
+
+@pytest.mark.parametrize("engine", ["voxcpm2", "omnivoice", "qwen_voice_design"])
+def test_request_contracts_preserve_all_new_engines(engine):
+    from backend.models import GenerationRequest, MCPClientBindingUpsert, SpeakRequest
+
+    assert GenerationRequest(profile_id="p", text="hello", engine=engine).engine == engine
+    assert SpeakRequest(text="hello", engine=engine).engine == engine
+    assert MCPClientBindingUpsert(client_id="test", default_engine=engine).default_engine == engine
+
+
+def test_voxcpm2_respects_force_cpu_override(monkeypatch):
+    monkeypatch.setenv("VOICEBOX_FORCE_CPU", "1")
+    assert VoxCPMBackend().device == "cpu"
+
+
+@pytest.mark.asyncio
+async def test_concurrent_model_loads_only_initialize_once(monkeypatch):
+    backend = VoxCPMBackend()
+    calls = []
+
+    def fake_load():
+        calls.append(True)
+        backend.model = object()
+
+    monkeypatch.setattr(backend, "_load_model_sync", fake_load)
+    await asyncio.gather(backend.load_model(), backend.load_model())
+    assert calls == [True]
diff --git a/docs/PROJECT_STATUS.md b/docs/PROJECT_STATUS.md
index ff15c087e..1c00658bd 100644
--- a/docs/PROJECT_STATUS.md
+++ b/docs/PROJECT_STATUS.md
@@ -23,14 +23,17 @@
 
 The backend exposes:
 
-- **`TTSBackend` Protocol** with seven concrete engine implementations:
+- **`TTSBackend` Protocol** with ten concrete engine implementations:
   - Qwen3-TTS (PyTorch or MLX depending on platform)
   - Qwen CustomVoice (predefined speakers with instruct)
+  - Qwen VoiceDesign (voices from written descriptions)
   - LuxTTS (fast, CPU-friendly)
   - Chatterbox Multilingual (23 languages)
   - Chatterbox Turbo (English, paralinguistic tags)
   - TADA (1B English, 3B multilingual via HumeAI)
   - Kokoro 82M (pre-built voices, CPU realtime)
+  - OmniVoice (23 exposed languages, non-commercial weights)
+  - VoxCPM2 (30-language local cloning)
 - **`STTBackend` Protocol** for Whisper (PyTorch or MLX-Whisper)
 - **Profiles / History / Stories** services for persistence and timeline editing
 
@@ -50,6 +53,9 @@ The backend exposes:
 | TADA | `backend/backends/hume_backend.py` | HumeAI TADA — 1B English + 3B Multilingual |
 | Kokoro | `backend/backends/kokoro_backend.py` | Kokoro 82M — CPU realtime, pre-built voices |
 | Qwen CustomVoice | `backend/backends/qwen_custom_voice_backend.py` | Qwen CustomVoice — predefined speakers with instruct |
+| Qwen VoiceDesign | `backend/backends/qwen_voice_design_backend.py` | Qwen VoiceDesign — voices from written descriptions |
+| OmniVoice | `backend/backends/omnivoice_backend.py` | OmniVoice — non-commercial multilingual cloning |
+| VoxCPM2 | `backend/backends/voxcpm_backend.py` | VoxCPM2 — 30-language local cloning |
 | Platform detect | `backend/platform_detect.py` | Apple Silicon → MLX, else → PyTorch |
 | API types | `backend/models.py` | Pydantic request/response models |
 | HF progress | `backend/utils/hf_progress.py` | HFProgressTracker (tqdm patching for download progress) |
@@ -69,7 +75,7 @@ The backend exposes:
 ```
 POST /generate
   1. Look up voice profile from DB
-  2. Resolve engine from request (qwen | qwen_custom_voice | luxtts | chatterbox | chatterbox_turbo | tada | kokoro)
+  2. Resolve engine from request or profile preference (qwen | qwen_custom_voice | qwen_voice_design | luxtts | chatterbox | chatterbox_turbo | tada | kokoro | omnivoice | voxcpm2)
   3. Get backend: get_tts_backend_for_engine(engine)  # thread-safe singleton per engine
   4. Check model cache → if missing, trigger background download, return HTTP 202
   5. Load model (lazy): tts_backend.load_model(model_size)
@@ -189,7 +195,7 @@ Shipped 2026-04-25 (PR #544). Voicebox went from a voice-cloning studio to a ful
 | Model | PR / Branch | Reason |
 |-------|-------------|--------|
 | **CosyVoice2/3** | PR #311 | Output quality too poor. Heavy deps, no PyPI, needed 5+ shims. PR should be closed. |
-| **VoxCPM 1.5 / VoxCPM2** | `voicebox-new-models` research (2026-04-18) | **Backlogged.** See detailed analysis below. |
+| **VoxCPM 1.5 / VoxCPM2** | `feat/voxcpm2-backend` | **Active community integration.** VoxCPM2 now has a native Voicebox backend draft using CUDA with macOS CPU fallback; see detailed analysis below. |
 
 #### VoxCPM — Evaluation Notes (2026-04-18)
 
@@ -203,18 +209,15 @@ Shipped 2026-04-25 (PR #544). Voicebox went from a voice-cloning studio to a ful
 - Zero-shot cloning + style control via parenthetical prefixes in text (`(slightly faster, cheerful tone)...`)
 - Relatively high-quality output per demos
 
-**Why we backlogged it:**
-- **Effectively CUDA-only.** README states `CUDA ≥ 12.0` as hard requirement. Source code's `from_pretrained(device=None|"auto")` claims "preferring CUDA, then MPS, then CPU," but in practice:
-  - **MPS (Apple Silicon) broken upstream** — OpenBMB/VoxCPM issues #232 (`NotImplementedError: Output channels > 65536 not supported at the MPS device`) and #248 (`IndexError` on M3 Mac) are both open with no resolution.
-  - **CPU unsupported in the Python package** — issue #256 shows `voxcpm --device cpu` rejected with `unrecognized arguments`. The only CPU path is the third-party **VoxCPM.cpp** GGML engine, which is a separate ecosystem project, not `pip install voxcpm`.
-  - **macOS source install fails** — issue #233 open with no resolution.
-- Would require CUDA-only gating in UI (new `requires_cuda` flag on `ModelConfig`, lock icon + "Requires NVIDIA GPU" in `ModelManagement.tsx` / `EngineModelSelector.tsx`) plus a hard error at `load_model()` as safety net. Doable but adds first-class platform gating that doesn't exist for any other engine today.
-- Voicebox's user base skews Apple Silicon (MLX is a primary backend). Shipping a CUDA-only model sets a precedent worth a separate scoping discussion (see issues #419 engine sprawl, #420 platform tiers, PR #465).
+**Earlier backlog assessment:**
+- The original 2026-04-18 assessment treated VoxCPM2 as effectively CUDA-only because CPU/MPS reports failed during model warm-up and generation.
+- Upstream subsequently added explicit device selection, safe MPS dtype handling, and a broadcastable SDPA mask for the CPU attention path.
+- Local validation on Apple Silicon with PyTorch 2.8.0 confirmed CPU loading, ordinary generation, reference-audio cloning, and a short MPS generation all complete successfully. MPS was substantially slower in the smoke test, so the integration follows existing Chatterbox/TADA precedent: CUDA where available and CPU on macOS.
 
-**What would change the decision:**
-- Upstream fixes MPS crashes (watch issues #232, #248).
-- We define an "experimental / CUDA-only" engine tier as part of issue #419 / PR #465, and decide it's acceptable to ship engines that are hidden on non-NVIDIA platforms.
-- VoxCPM.cpp matures into a viable CPU path we can wrap (currently separate project, C++/GGML, unclear ergonomics).
+**Remaining validation before shipping:**
+- Exercise the frozen PyInstaller binary with a clean model cache.
+- Validate CUDA generation on Windows/Linux hardware.
+- Keep macOS on CPU until MPS performance and stability are consistently better.
 
 **Integration shape if we revive it:** Zero-shot cloning maps naturally to the Chatterbox-style backend (store `ref_audio` + `ref_text` paths in the voice prompt dict, process at generate time). Est. ~250 lines for `voxcpm_backend.py` + one `ModelConfig` entry + engine registration in `backends/__init__.py`. Frontend UI gating is the bigger lift.
 
@@ -260,7 +263,7 @@ Shipped 2026-04-25 (PR #544). Voicebox went from a voice-cloning studio to a ful
 
 - **Thread-safe backend registry** (`_tts_backends` dict + `_tts_backends_lock`) with double-checked locking
 - **Per-engine backend instances** — each engine gets its own singleton, loaded lazily
-- **Engine field on GenerationRequest** — frontend sends `engine: 'qwen' | 'qwen_custom_voice' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo' | 'tada' | 'kokoro'`
+- **Engine field on GenerationRequest** — frontend sends `engine: 'qwen' | 'qwen_custom_voice' | 'qwen_voice_design' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo' | 'tada' | 'kokoro' | 'omnivoice' | 'voxcpm2'`
 - **Per-engine language filtering** — `ENGINE_LANGUAGES` map in frontend, backend regex accepts all languages
 - **Per-engine voice prompts** — `create_voice_prompt_for_profile()` dispatches to the correct backend
 - **Profile type system** — preset vs cloned profiles, UI grays out incompatible engines and auto-switches on selection
@@ -578,7 +581,7 @@ Notable:
 | **HumeAI TADA 1B/3B** | Zero-shot | 5x faster than LLM-TTS | 24 kHz | EN (1B), 10 (3B) | Medium | Partial — prosody | PyTorch | **Shipped** (PR #296) |
 | **Kokoro-82M** | Preset voices | CPU realtime | 24 kHz | 8 | Tiny (82M) | None | All | **Shipped** (PR #325) |
 | ~~**CosyVoice2-0.5B**~~ | 3-10s zero-shot | Very fast | 24 kHz | Multilingual | Low | **Yes** | — | **Abandoned** (PR #311) — poor output quality |
-| ~~**VoxCPM2**~~ | Zero-shot | ~0.15 RTF streaming | 48 kHz | 30 | Medium | Partial — parenthetical style | **CUDA-only in practice** | **Backlogged** (2026-04-18) — see notes above |
+| **VoxCPM2** | Zero-shot | ~0.15 RTF streaming on CUDA | 48 kHz | 30 | Medium | Partial — parenthetical style | CUDA + CPU fallback | **Integration in progress** — see notes above |
 | **Fish Speech** | 10-30s few-shot | Real-time | 24-44 kHz | 50+ | Medium | **Yes** — word-level inline | All | Candidate — license TBD |
 | **Fish Audio S2** | — | — | — | — | — | — | — | Candidate (#385) |
 | **XTTS-v2** | 6s zero-shot | Mid-GPU | 24 kHz | 17+ | Medium | Partial — style transfer from ref | All | Candidate — CPML license likely blocker |
@@ -687,7 +690,7 @@ The generation form now uses a flat model dropdown with engine-based routing. Pe
 
 ### 7. Engine Sprawl — NEW
 
-Eight TTS engines shipped, more candidates queued. Issue #419 asks for a first-class vs experimental distinction. Related: issue #420 asks for formalized platform support tiers. Combined, these would let us ship more engines more confidently with clearer expectations for users.
+Ten TTS engines are represented by the active integration set, with more candidates queued. Issue #419 asks for a first-class vs experimental distinction. Related: issue #420 asks for formalized platform support tiers. Combined, these would let us ship more engines more confidently with clearer expectations for users.
 
 ---
 
diff --git a/docs/content/docs/developer/model-management.mdx b/docs/content/docs/developer/model-management.mdx
index 6572484b1..591c13d15 100644
--- a/docs/content/docs/developer/model-management.mdx
+++ b/docs/content/docs/developer/model-management.mdx
@@ -7,7 +7,7 @@ description: "How model downloading, loading, and status tracking works across a
 
 Voicebox manages two categories of models:
 
-**TTS Models** — Eight engines covering zero-shot cloning, preset voices and voice design. Each engine may have one or more size variants.
+**TTS Models** — Ten engines covering zero-shot cloning, preset voices and voice design. Each engine may have one or more size variants.
 
 **ASR Models** — Whisper for transcription. Five sizes, plus MLX-Whisper on Apple Silicon for ~8× faster transcription.
 
@@ -28,9 +28,13 @@ Every model is described by a `ModelConfig` entry in `backend/backends/__init__.
 | **TADA 1B** | `tada` | `HumeAI/tada-1b` | 4 GB | ~4 GB | English |
 | **TADA 3B Multilingual** | `tada` | `HumeAI/tada-3b-ml` | 8 GB | ~8 GB | 10 |
 | **Kokoro 82M** | `kokoro` | `hexgrad/Kokoro-82M` | 350 MB | ~150 MB | 8 |
+| **VoxCPM2** | `voxcpm2` | `openbmb/VoxCPM2` | 4.8 GB | ~5 GB | 30 |
+| **OmniVoice (non-commercial)** | `omnivoice` | `k2-fsa/OmniVoice` | 3.3 GB | — | 23 exposed |
 
 On Apple Silicon, Qwen TTS uses MLX-optimized repos from `mlx-community` instead of the PyTorch repos. The backend picks automatically via `get_backend_type()`. CustomVoice and VoiceDesign are PyTorch-only — upstream publishes no MLX conversion.
 
+VoxCPM2 uses CPU on macOS and CUDA when available on other platforms. MPS is deliberately disabled for this backend.
+
 Qwen VoiceDesign ships a single 1.7B checkpoint; there is no 0.6B variant. It is the only engine that consumes `voice_type: "designed"` profiles, where the voice identity is the profile's `design_prompt` rather than reference audio or a preset id.
 
 ## Available Whisper Models
diff --git a/docs/content/docs/developer/setup.mdx b/docs/content/docs/developer/setup.mdx
index d4a777a52..f00f506c8 100644
--- a/docs/content/docs/developer/setup.mdx
+++ b/docs/content/docs/developer/setup.mdx
@@ -183,7 +183,7 @@ Models are automatically downloaded from HuggingFace Hub on first use, with live
 See [Model Management](/developer/model-management) for the full list.
 
 <Callout type="warn">
-  First-time usage will be slower due to model downloads, but subsequent runs will use cached models.
+  First-time usage will be slower due to model downloads, but subsequent runs will use cached models. VoxCPM2 requires roughly 4.8 GB and outputs 48kHz audio.
 </Callout>
 
 ## Generate OpenAPI Client
@@ -235,9 +235,12 @@ pip install -r requirements-mlx.txt
 # Chatterbox pins numpy<1.26 / torch==2.6 which break on Python 3.12+
 pip install --no-deps chatterbox-tts
 
-# HumeAI TADA pins torch>=2.7,<2.8 which conflicts with our torch>=2.1
+# HumeAI TADA pins a narrow torch range; keep Voicebox's shared torch build
 pip install --no-deps hume-tada
 
+# Voicebox uses VoxCPM2 synthesis without its optional UI/training/denoiser stack
+pip install --no-deps "voxcpm>=2.0.3,<2.1"
+
 # Install Qwen3-TTS from source
 pip install git+https://github.com/QwenLM/Qwen3-TTS.git
 
diff --git a/docs/content/docs/developer/tts-generation.mdx b/docs/content/docs/developer/tts-generation.mdx
index 4e2b9d0c6..e4a4d425b 100644
--- a/docs/content/docs/developer/tts-generation.mdx
+++ b/docs/content/docs/developer/tts-generation.mdx
@@ -142,7 +142,7 @@ The request path from frontend to audio file:
      "max_chunk_chars": 800
    }
    ```
-   The `engine` field is validated against the regex `^(qwen|qwen_custom_voice|luxtts|chatterbox|chatterbox_turbo|tada|kokoro)$`.
+   The `engine` field accepts `qwen`, `qwen_custom_voice`, `qwen_voice_design`, `luxtts`, `chatterbox`, `chatterbox_turbo`, `tada`, `kokoro`, `omnivoice`, or `voxcpm2`. When omitted, the profile's default engine is used.
 
 2. **Route** — `routes/generate.py` validates input and delegates.
 
diff --git a/docs/content/docs/index.mdx b/docs/content/docs/index.mdx
index e0c4f5d64..0be1be928 100644
--- a/docs/content/docs/index.mdx
+++ b/docs/content/docs/index.mdx
@@ -5,7 +5,7 @@ description: "Voicebox is the open-source, local-first AI voice studio — a fre
 
 Voicebox is the **open-source, local-first AI voice studio** — a free
 alternative to ElevenLabs and WisprFlow in one app. Clone voices, generate
-speech across 8 TTS engines, dictate into any app with a global hotkey,
+speech across 10 TTS engines, dictate into any app with a global hotkey,
 compose multi-voice projects, and let any MCP-aware agent speak in a voice
 you own. Everything runs on your hardware.
 
@@ -15,9 +15,9 @@ you own. Everything runs on your hardware.
 - **Captures tab** — paired audio + transcript archive, retranscribe / refine / play-as-voice
 - **Voice personalities** — per-profile compose button + persona-rewrite toggle, powered by a local LLM
 - **Agents speak back** — any MCP-aware agent can call Voicebox to speak in one of your cloned voices
-- **8 TTS engines** — Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro
+- **10 TTS engines** — Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro, OmniVoice (non-commercial), VoxCPM2
 - **Cloning and preset voices** — zero-shot cloning or 50+ curated preset voices
-- **23 languages** — from English to Arabic, Japanese, Hindi, Swahili
+- **30 languages** — from English to Arabic, Japanese, Hindi, Thai, Vietnamese
 - **Post-processing effects** — pitch shift, reverb, delay, chorus, compression, filters
 - **Expressive speech** — paralinguistic tags (`[laugh]`, `[sigh]`) and natural-language delivery control
 - **Unlimited length** — auto-chunking with crossfade for long scripts
diff --git a/docs/content/docs/overview/creating-voice-profiles.mdx b/docs/content/docs/overview/creating-voice-profiles.mdx
index 2813b5841..901fdfeb2 100644
--- a/docs/content/docs/overview/creating-voice-profiles.mdx
+++ b/docs/content/docs/overview/creating-voice-profiles.mdx
@@ -27,7 +27,7 @@ Use this when you want to replicate a specific person's voice from a recording.
     10-30 seconds of clear speech, minimal background noise. See [Voice Cloning](/overview/voice-cloning) for the engine catalog.
   </Step>
   <Step title="Create Profile">
-    **Profiles** → **+ New Profile** → choose a cloning engine (Qwen3-TTS, Chatterbox Multilingual, Chatterbox Turbo, LuxTTS, or TADA)
+    **Profiles** → **+ New Profile** → choose a cloning engine (Qwen3-TTS, Chatterbox Multilingual, Chatterbox Turbo, LuxTTS, TADA, or VoxCPM2)
   </Step>
   <Step title="Upload or Record Sample">
     Drag in an audio file, or record directly with the in-app recorder
diff --git a/docs/content/docs/overview/introduction.mdx b/docs/content/docs/overview/introduction.mdx
index 0a1f53a18..88b7541b1 100644
--- a/docs/content/docs/overview/introduction.mdx
+++ b/docs/content/docs/overview/introduction.mdx
@@ -27,7 +27,7 @@ shared between input and output.
   into the focused field on macOS and Windows (see [Dictation](/overview/dictation))
 - **Captures tab** — paired audio + transcript archive, retranscribe,
   refine, play-as-voice, promote-to-sample (see [Captures](/overview/captures))
-- **Voice cloning** — 5 cloning engines covering 23 languages. Zero-shot
+- **Voice cloning** — 7 cloning engines covering 30 languages. Zero-shot
   cloning from a reference sample (see [Voice Cloning](/overview/voice-cloning))
 - **Preset voices** — 50+ curated voices via Kokoro and Qwen CustomVoice
   for when you don't want to clone (see [Preset Voices](/overview/preset-voices))
@@ -46,7 +46,7 @@ shared between input and output.
 
 ## TTS Engines
 
-Eight engines with different strengths, switchable per-generation:
+Ten engines with different strengths, switchable per-generation:
 
 | Engine | Profile Type | Languages | Strengths |
 |--------|--------------|-----------|-----------|
@@ -58,6 +58,8 @@ Eight engines with different strengths, switchable per-generation:
 | **Chatterbox Turbo** | Cloned | English | Fast 350M model with paralinguistic emotion/sound tags |
 | **TADA** (1B / 3B) | Cloned | 10 | HumeAI speech-language model — 700s+ coherent audio |
 | **Kokoro** | Preset (50 voices) | 9 | 82M parameters, CPU realtime, lowest VRAM of any engine |
+| **VoxCPM2** | Cloned | 30 | 2B multilingual model, 48kHz output, transcript-guided cloning |
+| **OmniVoice** | Cloned | 23 exposed | Multilingual cloning and attribute instructions; non-commercial weights |
 
 ## STT and local LLM
 
@@ -101,7 +103,7 @@ No cloud fallback, no bring-your-own-API-key. Local is the product.
 | Frontend | React, TypeScript, Tailwind CSS |
 | State | Zustand, React Query |
 | Backend | FastAPI (Python) |
-| TTS Engines | Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox, Chatterbox Turbo, TADA, Kokoro |
+| TTS Engines | Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox, Chatterbox Turbo, TADA, Kokoro, OmniVoice (non-commercial), VoxCPM2 |
 | STT | Whisper / Whisper Turbo (PyTorch or MLX) |
 | Local LLM | Qwen3 0.6B / 1.7B / 4B (MLX or PyTorch) |
 | Effects | Pedalboard (Spotify) |
diff --git a/docs/content/docs/overview/mcp-server.mdx b/docs/content/docs/overview/mcp-server.mdx
index 4acdeee05..3d36c7645 100644
--- a/docs/content/docs/overview/mcp-server.mdx
+++ b/docs/content/docs/overview/mcp-server.mdx
@@ -116,7 +116,7 @@ for the shim to connect.
 voicebox_speak({
   text: "Deploy complete.",
   profile?: "Morgan",            // name or id; falls back to per-client binding, then default
-  engine?: "qwen",               // qwen | qwen_custom_voice | luxtts | chatterbox | chatterbox_turbo | tada | kokoro
+  engine?: "qwen",               // qwen | qwen_custom_voice | qwen_voice_design | luxtts | chatterbox | chatterbox_turbo | tada | kokoro | omnivoice | voxcpm2
   personality?: true,            // rewrite via the profile's personality LLM before TTS; default comes from the per-client binding
   language?: "en",
 })
diff --git a/docs/content/docs/overview/remote-mode.mdx b/docs/content/docs/overview/remote-mode.mdx
index 12abcfc63..7b3f9274a 100644
--- a/docs/content/docs/overview/remote-mode.mdx
+++ b/docs/content/docs/overview/remote-mode.mdx
@@ -35,6 +35,7 @@ In Remote Mode, the Voicebox desktop app (running on your local machine) communi
     # Engines with incompatible transitive pins — install with --no-deps
     pip install --no-deps chatterbox-tts
     pip install --no-deps hume-tada
+    pip install --no-deps "voxcpm>=2.0.3,<2.1"
 
     # Qwen3-TTS from source
     pip install git+https://github.com/QwenLM/Qwen3-TTS.git
diff --git a/docs/content/docs/overview/troubleshooting.mdx b/docs/content/docs/overview/troubleshooting.mdx
index f0e4ff4a4..fe8eb1d51 100644
--- a/docs/content/docs/overview/troubleshooting.mdx
+++ b/docs/content/docs/overview/troubleshooting.mdx
@@ -338,6 +338,7 @@ This is expected behavior. The first generation downloads the selected TTS engin
     pip install -r requirements.txt
     pip install --no-deps chatterbox-tts
     pip install --no-deps hume-tada
+    pip install --no-deps "voxcpm>=2.0.3,<2.1"
     pip install git+https://github.com/QwenLM/Qwen3-TTS.git
     ```
   </Accordion>
diff --git a/docs/content/docs/overview/voice-cloning.mdx b/docs/content/docs/overview/voice-cloning.mdx
index 400e114f5..ae6fc87b0 100644
--- a/docs/content/docs/overview/voice-cloning.mdx
+++ b/docs/content/docs/overview/voice-cloning.mdx
@@ -7,7 +7,7 @@ description: "Clone any voice from a few seconds of reference audio"
 
 Voicebox can replicate a specific person's voice from a short audio sample — known as **zero-shot voice cloning**. You provide 10-30 seconds of clear speech, the model extracts a voice embedding, and from then on you can generate any text in that voice.
 
-Five engines in 0.4 support cloning:
+Six engines support cloning:
 
 | Engine                      | Languages | Strengths                                                                  |
 | --------------------------- | --------- | -------------------------------------------------------------------------- |
@@ -16,6 +16,7 @@ Five engines in 0.4 support cloning:
 | **Chatterbox Turbo**        | English   | Fast 350M model with paralinguistic emotion tags (`[laugh]`, `[sigh]`)     |
 | **LuxTTS**                  | English   | Lightweight (~1 GB VRAM), 48 kHz output, 150x realtime on CPU              |
 | **TADA** (1B / 3B)          | 10        | Speech-language model with 700s+ coherent long-form generation             |
+| **VoxCPM2**                 | 30        | 2B multilingual model, 48 kHz output, transcript-guided high-fidelity cloning |
 
 <Callout type="info">
   Don't want to record audio? Use a curated voice from Kokoro or Qwen CustomVoice instead — see [Preset Voices](/overview/preset-voices).
@@ -50,6 +51,7 @@ Different engines suit different use cases. The profile grid greys out unsupport
 | Expressive English with `[laugh]` `[sigh]` tags    | **Chatterbox Turbo**  |
 | CPU-only or GPU-light setup, English               | **LuxTTS**            |
 | Long-form generation (audiobooks, full chapters)   | **TADA 3B**           |
+| Broad multilingual cloning with 48 kHz output      | **VoxCPM2**           |
 
 ## Best Practices
 
@@ -89,6 +91,7 @@ Adding multiple samples from the same speaker can improve quality:
 - **Chatterbox Turbo** — English
 - **LuxTTS** — English
 - **TADA 3B** — 10 multilingual; **TADA 1B** — English
+- **VoxCPM2** — 30 languages; CUDA acceleration with CPU fallback
 
 For complete language tables and engine-specific notes, see the [TTS Engines developer guide](/developer/tts-engines).
 
diff --git a/justfile b/justfile
index 4e12685a8..7e8ad04ce 100644
--- a/justfile
+++ b/justfile
@@ -125,6 +125,8 @@ setup-python:
     # OmniVoice requires transformers>=5.3; we vendor the one class it
     # needs from 5.x (see backend/vendor/higgs_audio_v2_tokenizer).
     "{{ python }}" -m pip install --no-deps omnivoice
+    # VoxCPM's optional UI/training/denoiser dependency tree is not used by Voicebox.
+    "{{ python }}" -m pip install --no-deps 'voxcpm>=2.0.3,<2.1'
     # Apple Silicon: install MLX backend
     if [ "$(uname -m)" = "arm64" ] && [ "$(uname)" = "Darwin" ]; then
         echo "Detected Apple Silicon — installing MLX dependencies..."
@@ -211,6 +213,7 @@ setup-python:
     Invoke-Pip install --no-deps chatterbox-tts; \
     Invoke-Pip install --no-deps hume-tada; \
     Invoke-Pip install --no-deps omnivoice; \
+    Invoke-Pip install --no-deps "voxcpm>=2.0.3,<2.1"; \
     Invoke-Pip install git+https://github.com/QwenLM/Qwen3-TTS.git; \
     Invoke-Pip install pyinstaller ruff pytest pytest-asyncio openai -q; \
     Write-Host "Python environment ready."
diff --git a/scripts/engine-capabilities.test.ts b/scripts/engine-capabilities.test.ts
new file mode 100644
index 000000000..39e0dd5e4
--- /dev/null
+++ b/scripts/engine-capabilities.test.ts
@@ -0,0 +1,30 @@
+import assert from 'node:assert/strict';
+import { test } from 'node:test';
+import { engineSupportsInstruct } from '../app/src/lib/constants/engines';
+import { getLanguageOptionsForEngine } from '../app/src/lib/constants/languages';
+
+test('instruction controls preserve upstream engines alongside VoxCPM2', () => {
+  for (const engine of ['qwen_custom_voice', 'qwen_voice_design', 'omnivoice', 'voxcpm2']) {
+    assert.equal(engineSupportsInstruct(engine), true, engine);
+  }
+  for (const engine of ['qwen', 'kokoro', undefined]) {
+    assert.equal(engineSupportsInstruct(engine), false, engine);
+  }
+});
+
+test('language menus retain separate engine capabilities', () => {
+  const voxcpm = getLanguageOptionsForEngine('voxcpm2');
+  const omni = getLanguageOptionsForEngine('omnivoice');
+  const design = getLanguageOptionsForEngine('qwen_voice_design');
+
+  assert.equal(voxcpm.length, 30);
+  assert.equal(new Set(voxcpm.map(({ value }) => value)).size, 30);
+  assert.ok(voxcpm.every(({ label }) => label));
+  assert.ok(voxcpm.some(({ value }) => value === 'th'));
+  assert.equal(omni.length, 23);
+  assert.equal(
+    omni.some(({ value }) => value === 'th'),
+    false,
+  );
+  assert.equal(design.length, 10);
+});