mistral.rs is a Rust-based LLM inference engine and serving platform built by Eric Buehler that focuses on speed and broad hardware/model coverage rather than tying users to one vendor's model family. It automatically detects a supported Hugging Face model's architecture, weight format, and chat template, and can also load local GGUF files directly, selecting a matching quantized artifact with a single --quant flag or falling back to its own UQFF quantization format when a prebuilt GGUF isn't available. Beyond text, the engine handles true multimodal workloads in one process -- vision, video, and audio input, plus speech and image generation and embeddings -- and exposes both OpenAI-compatible /v1 endpoints and Anthropic-compatible Messages endpoints from the same mistralrs serve process, alongside Prometheus metrics and a built-in web UI at /ui for inspecting reasoning traces, code execution, and plots. A native agentic runtime adds web search, local Python and shell execution, OpenAI-compatible Skills, session management, and custom tool hooks, usable directly from an mistralrs run --agent REPL. An included mistralrs tune command recommends quantization and device mapping based on the detected model and hardware. The project publishes per-release CUDA benchmarks that document commands, model revisions, and host metadata, comparing its UQFF Q8 throughput against llama.cpp GGUF Q8_0 and its BF16 throughput against vLLM on GB10, B200, and H100 SXM. The published tables favor mistral.rs on the quantized runs while trailing vLLM on several BF16 prefill rows. It ships prebuilt binaries for Linux, macOS, and Windows via a one-line install script, along with Python and Rust SDKs for embedding the engine directly into applications.