Fermion Research
The lab behind neutrino-1, released 2026-07-27. Discovered while ingesting that announcement, which is the only source this wiki holds on it.
What its own site establishes. It ships models rather than an API: three open-weight models
under Apache-2.0 with no waitlist, an engine (pip install fermion-research), a public llama.cpp
fork, and an MLX pack. It publishes at unusual depth — the Neutrino-1 post is accompanied by two
further research posts the same day, one on native-format training and one on the engine — while
keeping the two things that matter commercially closed: the format’s “proprietary structure” and
the training method. The site carries Models, Research, Newsroom and Careers sections and a
trademark symbol, so it presents as a company rather than a project.
What the wiki does not know. No founders, no funding, no headcount, no location, no publication history, no named authors on the release. No independent evaluation of its models exists. No pricing, because there is no service to price. Its business model is not stated anywhere on the page.
The open question. The Neutrino-1 architecture matches qwen‘s published Qwen3-8B and Qwen3-0.6B configurations field for field, and the release never mentions it — recorded on neutrino-1 with both readings and no verdict. Whether this lab designs models or reformats other people’s is what one release cannot settle.
Connections
- publisher of neutrino-1 and ternary-weights
- architecture matches qwen‘s Qwen3 family