Machine learning infrastructure in Rust, from tensors to inference servers. The stack needs no Python runtime.
| Project | What it does |
|---|---|
| blazr | Inference server for LLMs, speech, and embeddings. OpenAI and Anthropic compatible API. Runs on CPU, CUDA, and WebGPU. C API for apps. |
| oxidizr | LLM training for Mamba, MLA, MoE, and hybrid models. |
| voxr | Audio and speech on boostr: decoding, TTS, ASR, and VAD. |
| compressr | Model conversion and quantization between SafeTensors and GGUF. |
| boostr | ML framework on numr: attention, MoE, Mamba, quantization, model loaders. |
| numr | Tensors, linear algebra, and FFT on CPU, CUDA, and WebGPU. |
| solvr | Scientific computing on numr: optimization, ODE and PDE solvers, statistics. |
| splintr | BPE, WordPiece, and SentencePiece tokenizers. Python bindings. |
| regexr | Regex engine with lookarounds and JIT, built for tokenization. |
| nexar | Distributed runtime with QUIC transport and collectives. |
| treeboost | Gradient boosted decision trees for large tabular data. |
| fluxbench | Benchmarking with crash isolation and bootstrap statistics. |
| venus | Reactive notebooks for Rust, stored as plain .rs files. |
| axp | Open protocol that gives AI agents a sandbox to run code and tools. |
| cargo-reclaim | Background companion to cargo clean for large Rust projects. |
numr
├── solvr
└── boostr
├── blazr inference
├── oxidizr training
├── voxr speech
└── compressr conversion
splintr, regexr tokenization, used by boostr and blazr
cargo install blazr --locked
blazr pull Qwen/Qwen2.5-0.5B
blazr list
blazr serve --model <name>See the blazr README for models, formats, and the API.
