Open code models trained only on Rust that compiled and passed its tests. Small enough to run on your own machine.
Open weights under Apache 2.0. Safetensors for Transformers and vLLM, GGUF for llama.cpp and Ollama.
Our main Rust model. Strong on everyday library code: traits, generics, iterators, error handling, tests.
A tiny Rust model for quick completions and fast local loops. Runs on a laptop CPU.
The largest Lin-1 model, for harder multi-step Rust tasks.
Rust versions of HumanEval and MBPP. Generated code is compiled and run against the reference tests.
| Model | Size | HumanEval p@1 / p@10 | MBPP p@1 / p@10 | Holdout | LBPP |
|---|
Holdout: our private set of 1,726 Rust tasks with tests that no model was trained on. LBPP: Rust split of Less Basic Python Problems (129 tasks). Holdout and LBPP were run only for Lindot and Temper models.
Small models copy what they see, so everything they see has to work.
Permissively licensed Rust from Stack-Edu and Strandset-Rust. Kept only if it compiles and passes its own unit tests.
Every sample is checked for 12-gram overlap with all benchmarks we report. Matches are dropped.
Reinforcement learning with one reward: the code compiles and hidden tests pass. Used for Epsilon Preview.
vllm serve LindotAI/Lin-1-Locus --max-model-len 8192llama-cli -hf LindotAI/Lin-1-Locus -p "Write a Rust LRU cache with tests"