Fermion Research

Every machine this lands on.

Neutrino-1 8B is a 2.56 GB download that decodes from one artifact on a datacenter GPU, a 16 GB MacBook, and a desktop CPU. This page collects the runs: the hardware, the rate, and the session.

The engine picks its backend at load and memory-maps the container as it arrives, so a submission needs nothing beyond the two commands and the machine they ran on.

A dark moss filament structure growing through a soft material field

Every showcase starts as a small machine on a bench.

Start here

Two commands, then the transcript.

The wheel pulls the engine and the weights together, so the pair always match. First-run behaviour, the digest gate, and the GGUF and MLX paths are documented in full.

Read the documentation
$ pip install fermion-research
$ fermion chat --model fermionresearch/Neutrino-8B

Two commands from a fresh machine to a streaming Neutrino-1 8B chat.

Submissions

Six categories are open.

One for each kind of run the engine was built to survive. A submission that fits none of them is still the sixth.

  1. 01

    First run

    The session from a machine that had never held a model. Two commands from a fresh install to a streaming reply, and whatever gets asked first.

  2. 02

    Pocket hardware

    Neutrino-1 0.6B is a 238 MB download that decodes at 225 to 236 tok/s on nine Apple M5 CPU threads. Send the smallest board that holds a conversation, with its rate.

  3. 03

    Off the grid

    One file, no network after the download, no server-side component in the loop. Cabins, boats, field stations, aircraft.

  4. 04

    Workbench build

    The engine inside a tool or an appliance: the enclosure, the power budget, and what the model is wired to do in it.

  5. 05

    Rate report

    The tokens-per-second line at the end of a session, with the machine stated exactly: chip, memory, thread count, and which of the three runtimes served it.

  6. 06

    Something else

    Runs that do not fit the five above. This is the category most likely to move the others.

Send a run

The session and the machine.

Plain email, no form. Runs are published with whatever name and link arrive with them, and the hardware line is printed exactly as written.

community@fermionresearch.com
What to send
  • The scrollback or a screenshot of the session
  • The chip, the memory, and the runtime that served it
  • One line about where the machine lives

The engine prints its backend and thread count at load; that line covers the hardware field.

Newsletter

One short email as the first runs go up.

Subscribe in the newsroom