Docs
Browse documentation

Docs / Reference

Model reference

The three Neutrino repositories, their runnable artifacts, and the integration each file is intended for.

Neutrino-1 8B

The default general model and the only Neutrino-1 model that reliably emits tool calls. The trained context window is 40,960 positions.

ArtifactSizeUsed by
neutrino-8b_v4.bin3.88 GBfermion native and PyTorch paths, plus MLX
neutrino-8b_v4.tv4z2.56 GBApple Silicon download transport
neutrino-8b-fv5.gguf4.09 GBFermion llama.cpp fork
bin/fermion-run-*about 2 MB eachOptimized native runtime
MANIFEST.json and SHA sidecarssmallIntegrity checks
Open Neutrino-8B on Hugging Face

Neutrino-1 0.6B

A 328 MB small model and the canonical --draft auto target. It can also run independently through chat, generate, serve, the native runners, or the model repository’s MLX pack.

  • Runnable container: 328 MB
  • Coded transport: 238 MB
  • GGUF: 343 MB
Open Neutrino-0.6B on Hugging Face

Neutrino-1 0.6B-Chat

A 328 MB conversational small model. It ships native runners and MLX code. It does not currently ship GGUF weights and should not be selected for reliable tool calling.

Open Neutrino-0.6B-Chat on Hugging Face

License and base model

The package and model releases use Apache-2.0. Neutrino-1 8B is derived from Qwen3-8B, which uses the same license.