FuriosaAI develops data center AI accelerators. Our RNGD (pronounced "Renegade") accelerator is designed for high-performance inference for LLMs and agentic AI.

Run text-generation, vision-language, embedding, and reranking models on RNGD with Furiosa-LLM. The featured repositories below include model files and a matching Furiosa Executable Bundle (FXB) containing pre-compiled NPU kernels.

Furiosa SDK 2026.4.0 adds Gemma 4, EXAONE 4.5 with experimental MTP support, Qwen3-VL 2B/4B variants, Mistral NeMo, and dynamically quantized FP8 Llama variants. Embedding and reranking coverage expands with BGE-M3, E5-Mistral, BGE Reranker v2 M3, Harrier OSS v1, and additional Qwen3 Embedding and Reranker sizes. See the 2026.4.0 Release Notes for details.

Follow the Quick Start to get set up. Once installed, furiosa-llm serve <repo> discovers a compatible FXB automatically. Check each model card for hardware requirements and model-specific launch options.

Need a model with custom configurations? Build your own FXB with fxb build. See the FXB guide for how bundles are built, cached, and distributed.

Visit Supported Models for architecture support and validation status, and learn more about RNGD at https://furiosa.ai/rngd.

Featured Pre-compiled Models

Browse all repositories at https://huggingface.co/furiosa-ai/models and curated sets at https://huggingface.co/furiosa-ai/collections.

The examples below target SDK 2026.4.0. When using an earlier SDK, select a compatible model revision and FXB.

Text Generation

Vision-Language Models

These models support text and image inputs. See the Vision-Language Models guide for serving and request examples.

Embedding Models

Reranking Models

Examples

First, install the prerequisites by following Installing Furiosa-LLM.

Then, start the Furiosa-LLM server. This example uses furiosa-ai/Qwen3-4B-FP8, a compact generative model:

furiosa-llm serve furiosa-ai/Qwen3-4B-FP8

Qwen3 is a hybrid reasoning model that thinks by default. To separate reasoning from the final response, launch it with the qwen3 reasoning parser:

furiosa-llm serve furiosa-ai/Qwen3-4B-FP8 \
    --reasoning-parser qwen3

Once the server is ready, send a request through the OpenAI-compatible Chat Completions API:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "furiosa-ai/Qwen3-4B-FP8",
        "messages": [
            {"role": "user", "content": "What is the capital of France?"}
        ]
    }' \
    | python -m json.tool

For image inputs, embeddings, and reranking, follow the examples in the corresponding model cards.

Learn more in the Quick Start with Furiosa-LLM, the per-model guides under Supported Models, and the FXB guide.