mistral.rs - Blazingly fast LLM inference.

mistral.rs - Fast, flexible LLM inference.

Latest

Muse Glimmer 30B: native text, image, and video inference with ATEM tool calling, reasoning controls, LoRA, ISQ/UQFF, and companion-projector GGUF loading. Model notes
GGUF loading: load a local file with -f, or select a published artifact with --quant. Tokenizer, configuration, and multimodal projector files are discovered when the available metadata identifies them unambiguously. Guide
OpenAI-compatible Skills: upload /v1/skills bundles and reference them from Responses requests for reusable procedures, helper scripts, and local data. Guide
OpenAI-compatible file inputs: upload /v1/files, attach Responses input_file or Chat file parts, and mount request files into shell/code sessions. Guide
DiffusionGemma: block-diffusion text generation. Fully integrated: paged attention, prefix caching, ISQ, multimodal, and tool calling. Guide
Anthropic Messages API: mistralrs serve now exposes Anthropic-compatible /v1/messages and /v1/messages/count_tokens endpoints alongside the OpenAI-compatible /v1 API. Guide
v0.8.2 CUDA performance: paged-attention and MoE optimizations deliver strong results on GB10, B200, and H100 SXM. Benchmarks
Agentic runtime: web search, local Python code execution, shell execution, OpenAI-compatible Skills, session management, and custom tool hooks. Guide
Gemma 4: full multimodal: text, image, video, and audio input. Supported models | Video setup

Benchmarks

v0.8.2 CUDA benchmarks

Mean tokens per second across prompt lengths and decode depths from 128 to 16384 tokens. Decode uses 256 generated tokens. See the full v0.8.2 report for commands, model revisions, host metadata, and appendix tables.

Q8 prefill TPS: mistral.rs UQFF q8 vs llama.cpp GGUF Q8_0

Model	Hardware	mistral.rs	llama.cpp
Gemma 4 E4B	GB10	7395.7	3973.7
Gemma 4 E4B	B200	27705.6	11992.4
Gemma 4 E4B	H100 SXM	26220.6	11702.1
Gemma 4 26B-A4B	GB10	2947.0	2178.5
Gemma 4 26B-A4B	B200	12725.3	8503.4
Gemma 4 26B-A4B	H100 SXM	12362.3	8055.1

Q8 decode TPS: mistral.rs UQFF q8 vs llama.cpp GGUF Q8_0

Model	Hardware	mistral.rs	llama.cpp
Gemma 4 E4B	GB10	44.1	40.5
Gemma 4 E4B	B200	241.4	194.4
Gemma 4 E4B	H100 SXM	223.1	183.0
Gemma 4 26B-A4B	GB10	46.8	46.4
Gemma 4 26B-A4B	B200	210.9	192.2
Gemma 4 26B-A4B	H100 SXM	199.8	183.9

BF16 prefill TPS: mistral.rs BF16 vs vLLM BF16

Model	Hardware	mistral.rs	vLLM
Gemma 4 E4B	GB10	5838.9	5812.9
Gemma 4 E4B	B200	43547.8	39431.2
Gemma 4 E4B	H100 SXM	35852.2	39293.7
Gemma 4 26B-A4B	GB10	592.2	3878.6
Gemma 4 26B-A4B	B200	3467.3	28532.8
Gemma 4 26B-A4B	H100 SXM	2766.0	26295.9

BF16 decode TPS: mistral.rs BF16 vs vLLM BF16

Model	Hardware	mistral.rs	vLLM
Gemma 4 E4B	GB10	25.1	18.8
Gemma 4 E4B	B200	202.6	196.2
Gemma 4 E4B	H100 SXM	174.4	153.0
Gemma 4 26B-A4B	GB10	26.9	23.2
Gemma 4 26B-A4B	B200	159.6	220.2
Gemma 4 26B-A4B	H100 SXM	138.7	148.0

Why mistral.rs?

Automatic model loading: Architecture, weight format, and chat template are detected for supported Hugging Face models and GGUF files, with flags available for explicit selection.
True multimodality: Text, vision, video, and audio, speech generation, image generation, and embeddings in one engine.
Quantization selection: --quant selects a matching artifact from GGUF repositories. For other Hugging Face repositories, it uses a prebuilt UQFF when available and otherwise applies ISQ. Docs
OpenAI + Anthropic compatible serving: The same mistralrs serve process exposes OpenAI-compatible /v1 endpoints and Anthropic-compatible Messages endpoints.
Prometheus metrics: mistralrs serve exposes a /metrics endpoint in Prometheus format, recording per-request counts and latency labeled by method, route, and status. Docs
Built-in web UI: Served at /ui by default. Shows reasoning, code execution, plots, and files inline. Edit any message and the new branch runs with its own Python state. Pass --no-ui to disable.
Hardware-aware: mistralrs tune recommends quantization and device mapping from the model config and your detected hardware.
Flexible SDKs: Python package and Rust crate to build your projects.
Native agentic support: built-in agentic loop with web search, local Python code execution, shell execution, OpenAI-compatible Skills, session management, and custom tool hooks.

Quick Start

Install

Linux/macOS:

curl -fsSL https://mistralrs.dev/install.sh | sh

Windows (PowerShell):

irm https://mistralrs.dev/install.ps1 | iex

Downloads a self-contained prebuilt binary for your platform (Metal on Apple Silicon; per-GPU CUDA or CPU on Linux; CPU on Windows), falling back to a source build if none matches. Standard acceleration needs no Rust or CUDA toolkit. Optional cuTile acceleration requires NVIDIA's separately installed tileiras tool.

Manual installation, accelerator details & other platforms

Run Your First Model

# Interactive chat
mistralrs run -m Qwen/Qwen3-4B

# One-shot prompt (no interactive session)
mistralrs run -m Qwen/Qwen3-4B -i "What is the capital of France?"

# One-shot with an image
mistralrs run -m google/gemma-4-E4B-it --image photo.jpg -i "Describe this image"

# Run a local GGUF or select a published 4-bit GGUF
mistralrs run -f /path/to/model.gguf
mistralrs run -m unsloth/Qwen3.5-4B-GGUF --quant 4

# Agentic REPL: search + code execution + shell from the terminal
mistralrs run --agent -m Qwen/Qwen3-4B

# Start an API server with the built-in web UI
mistralrs serve -m google/gemma-4-E4B-it

For the server command, visit http://localhost:1234/ui for the web chat interface. OpenAI-compatible clients use http://localhost:1234/v1; Anthropic-compatible clients use http://localhost:1234.

The `mistralrs` CLI

The CLI uses the same run, serve, and bench commands for model repositories, local directories, and GGUF files.

Auto-detection: Automatically detects model architecture, quantization format, and chat template
All-in-one: Single binary for chat, server, benchmarks, and web UI (run, serve, bench)
Hardware-aware tuning: mistralrs tune recommends quantization and device mapping for your model and hardware
Model formats: Hugging Face checkpoints, GGUF files, and UQFF quantizations

# Recommend settings for your hardware and emit a config file
mistralrs tune -m Qwen/Qwen3-4B --emit-config config.toml

# Run using the generated config
mistralrs from-config -f config.toml

# Diagnose system issues (CUDA, Metal, Hugging Face connectivity)
mistralrs doctor

Full CLI documentation

UI Demo

What Makes It Fast

Performance

Continuous batching support by default on all devices.
CUDA with FlashAttention V2/V3, Metal, and multi-GPU/distributed inference
PagedAttention for high throughput continuous batching on CUDA or Apple Silicon, prefix caching (including multimodal)

Quantization (full docs)

In-situ quantization (ISQ) for Hugging Face models
GGUF (2-8 bit), GPTQ, AWQ, HQQ, FP8, BNB support
⭐ Per-layer topology: Fine-tune quantization per layer for optimal quality/speed
⭐ Auto-select fastest quant method for your hardware

Flexibility

LoRA & X-LoRA with per-request LoRA selection and X-LoRA adapter mixing
AnyMoE: Create mixture-of-experts on any base model
Multiple models: Load/unload at runtime

Agentic Features

Integrated tool calling with grammar enforcement and strict schema mode
⭐ Server-side agentic loop: auto-execute tools and feed results back
⭐ Python code execution: persistent Jupyter-like sessions with matplotlib capture and multimodal feedback
⭐ Shell execution: persistent command-line sessions with sandboxing and approval controls
⭐ OpenAI-compatible Skills: uploaded skill bundles for Responses API agents
⭐ OpenAI-compatible file inputs: /v1/files, Responses input_file, Chat file, and workdir mounts
⭐ Web search integration with embedding-based ranking
⭐ Tool dispatch URL: POST tool calls to your own endpoint
⭐ MCP client: Connect to external tools via Process, HTTP, or WebSocket
Python/Rust tool callbacks for custom execution

Full feature documentation

Supported Models

Text, multimodal, speech, image generation, and embedding models across 45+ architectures. The supported models reference is the single source of truth: it explains how to check whether your model's config.json is supported, lists every architecture with copy-paste run commands, and is generated directly from the engine's loader registry so it never drifts.

Supported models reference | Request a new model

Python SDK

pip install mistralrs

In-process inference from Python: load a model with Runner and send OpenAI-shaped requests, no server required. Accelerator-specific wheels (CUDA, Metal, MKL, Accelerate) are listed in the getting-started guide.

Get started | API reference | Examples

Rust SDK

cargo add mistralrs

Embed the engine in a Rust application with the high-level mistralrs crate.

Get started | docs.rs | Crate | Examples

Docker

Prebuilt CPU and CUDA images are published to GHCR. Pull commands, tags, and Kubernetes notes are in the Docker guide.

Documentation

For complete documentation, see the Documentation.

Quick Links:

Quickstart - Install, first run, first serve
CLI Reference - All commands and options
Anthropic Messages API - Anthropic-compatible Messages, streaming, tool use, and token counting
HTTP API - OpenAI-compatible and Anthropic-compatible endpoints
Quantization - ISQ, GGUF, GPTQ, and more
Multi-GPU and Distributed - NCCL TP, P2P layer mapping, multi-node, and ring
MCP Integration - MCP integration documentation
Troubleshooting - Common issues and solutions
Environment variables - Environment variables for configuration

Citation

If you use mistral.rs in your research, please cite:

@misc{mistralrs,
  author = {Buehler, Eric},
  title = {{mistral.rs}: Fast, flexible {LLM} inference},
  year = {2024},
  url = {https://github.com/EricLBuehler/mistral.rs}
}

Citation metadata is available in CITATION.cff.

Contributing

Contributions welcome! Please open an issue to discuss new features or report bugs. If you want to add a new model, please contact us via an issue and we can coordinate.

Credits

This project would not be possible without the excellent work at Candle. Thank you to all contributors!

mistral.rs is not affiliated with Mistral AI.