500+ models · Compile once, run anywhere · AI model serving on RBLN NPUs

RBLN Model Zoo

---

Quick Start

The rbln-zoo CLI discovers models; each model is run from its own directory.

1. Discover — install the CLI and browse the catalog:

git clone https://github.com/RBLN-SW/rbln-model-zoo.git && cd rbln-model-zoo
uv pip install -e .

rbln-zoo list -s llama      # search the catalog
rbln-zoo cards              # show card types

2. Run — from the model's directory, install its dependencies and execute:

cd Rebel100/huggingface/transformers/text2text-generation/llama/llama3.1-8b
uv pip install -r requirements.txt
python compile.py && python inference.py     # single-file examples: python main.py

Every example lives under the product it targets — Rebel100/ for Rebel100™ (RBLN-CR13), ATOM/ for ATOM™+ and ATOM™-Max. A Rebel100/ example is its ATOM/ counterpart retuned for a single NPU, so swapping the leading directory is usually all it takes to move between them.

Important

Compilation requires the RBLN Compiler from RBLN's private package index. See the installation guide.

---

CLI

rbln-zoo browses and filters the model catalog; it does not compile or run models.

rbln-zoo list -c RBLN-CR13 -t text2text-generation -s qwen   # filter by card, task, keyword
rbln-zoo cards                                               # card types and counts

Command	Description	Flags
list	Browse and filter models	-c card · -f framework · -t task · -s search · -r refresh
cards	Show card types and counts	-r refresh
config	Read or write persistent CLI defaults	subcommands set · get · unset · list · path

list prints a table, which elides long paths. --paths prints one full model path per line instead — for a shell pipeline, or to copy — and --json prints the full entries.

The catalog is cached against the current commit and model_registry.yaml, so it refreshes itself once a change is committed. Pass -r to pick up a model that is not.

Card types

Models are tagged with RBLN product cards — RBLN-CR13 (Rebel100™), RBLN-CA22 (ATOM™+) and RBLN-CA25 (ATOM™-Max) — per the version matrix. Each product tree carries its cards by default: models under Rebel100/ are tagged RBLN-CR13, models under ATOM/ are tagged RBLN-CA22/RBLN-CA25. Matching is case-insensitive and honors aliases declared in model_registry.yaml.

Example — adding a card with aliases

cards:
  RBLN-CA22:
    description: "ATOM™+"
  RBLN-CA25:
    description: "ATOM™-Max"
  RBLN-CR13:
    description: "Rebel100™"
  CX:
    description: "Next-gen NPU"
    aliases: [RBLN-CX01]   # -c RBLN-CX01 resolves to CX

default_cards:            # per product tree (top-level directory)
  ATOM: [RBLN-CA22, RBLN-CA25]
  Rebel100: [RBLN-CR13]

overrides:
  ATOM/huggingface/transformers/.../model-a:
    cards: [RBLN-CA25, CX]

---

Ecosystems

Ecosystem	Models	Key packages
Hugging Face	150+	transformers, diffusers
PyTorch	250+	torch
TensorFlow	75+	keras, tensorflow

Note

Model counts are approximate, as of 2026-07-13 — see the Model Zoo for the live catalog.

C API — C/C++ inference bindings; install via APT, then build from source.

---

Deployment

Compile a model, then serve it on a supported inference server.

vLLM RBLN

A single vllm serve compiles the model on first run and serves the OpenAI-compatible API — no separate compile.py step.

uv pip install \
  --extra-index-url https://wheels.vllm.ai/0.30.0/cpu \
  --torch-backend cpu \
  vllm-rbln

On Rebel100™, one NPU serves Llama-3.1-8B-Instruct:

vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --model-impl transformers \
  --rbln-num-devices-per-local-rank 1 \
  --served-model-name Llama-3.1-8B-Instruct \
  --max-model-len 32768 \
  --block-size 8192 \
  --max-num-seqs 1

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Llama-3.1-8B-Instruct",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_tokens": 32
  }'

--rbln-num-devices-per-local-rank is how many NPUs the model is partitioned across. ATOM™ runs the same model across 8 with a 131,072-token window (--max-model-len 131072 --block-size 16384).

Note

Per-model serve commands live in the vLLM RBLN recipes — see them for the engine arguments a given model needs.

Other serving options

• NVIDIA Triton Inference Server — multi-model inference
• TorchServe — PyTorch model serving

---

Links

• CHANGELOG — release history
• Issues — report bugs, request features or new models