Recipes
One file says what to train: model, method and data. Run it, resume it, fork it, sweep it. Every run is frozen into a lock you can reproduce.
Local LLM development
Train, fine-tune, serve and connect language models on your own hardware, from one recipe. No cluster, no lock-in.
Early development Follow Tensorun-AI on GitHub
One file says what to train: model, method and data. Run it, resume it, fork it, sweep it. Every run is frozen into a lock you can reproduce.
LoRA, DPO, ORPO and GRPO on pretrained Qwen2, Llama and Mistral models, with a planner that fits the job to your GPU and a memory guard that recovers from out-of-memory.
One command serves a run, a model folder or a GGUF file behind an OpenAI-style /v1 API and an Ollama-style /api.
Point your coding agent or editor at your own model with ready-made profiles for pi, Zed and more.
Docs, downloads and benchmarks will live here once Tensorun is ready for you to try.