Fermion Research
An oil painting of bubble-chamber particle tracks, sage-green spirals and one lime track on a bright cream ground

Foundation models at
the floor of arithmetic.

AI research and models that push intelligence to the physical limits of computation.

Discover the models

Announcement

Introducing the Neutrino-1 models

Read the release

Mission

Intelligence is becoming the most valuable computation on earth, and the most expensive. Fermion Research exists to change the second part. We build foundation models that push intelligence to the physical limits of computation.

Nature runs a general intelligence on twenty watts. We do not believe that is the floor.

The work

Intelligence is leaving the datacenter.

A cobalt cellular membrane becoming progressively finer

Capability survives below floating point.

Neutrino
I'm on a flight with no wifi. Can you still help me outline the pitch?
Of course. I run right here on your laptop, connected or not. Give me the one sentence your pitch has to land and we will build backwards from it.
Okay: small models can be frontier models.
Good spine. Open with the claim, prove it with a demo instead of a chart, and close on what it makes possible. Want me to draft the opening thirty seconds?
Message Neutrino↑

Neutrino-1 8B, answering locally.

A 2.56 GB download, decoding on the machine in front of you.

A precise sandstone arc embedded in a tactile material field

The model, the format, and the engine are one act of design.

We train models natively in the format they ship in and build the engine that serves them. Nothing between the weights and the silicon is someone else's work.

Showing item 1 of 4: An amber cellular orbit representing a model on a laptop
A moss filament structure representing a model in a sensor
An amber cellular orbit representing a model on a laptop
A cobalt cellular membrane representing a model on a phone
A compact sandstone arc representing a model in a badge

Small enough fora laptop.

Intelligence as a property of things, the way software is.

From the lab

All updates
An amber cellular orbit gathering around a compact central form

Announcement

Introducing the Neutrino-1 models

Three models in a proprietary ternary-family format, served by one engine from one artifact on datacenter GPUs, Apple silicon, and desktop CPUs.

Read the announcement
A cobalt cellular membrane becoming progressively finer

Research

Intelligence at one-eighth the bits

Native-format training holds frontier capability in a ternary-family format at one eighth the bits of fp16; rounding a trained model to the same depth lands near chance.

A honey-toned filament carrying energy across a tactile surface

Research

The Neutrino Engine

The runtime built alongside the format: single-stream decode from one artifact on every backend, with a drafted path verified token for token against plain decode.

Models

The Neutrino family

Neutrino-1 8B against floating point

The same eight billion parameters in four storage forms, and the decode rate of the shipping artifact on four platforms.

Storage

Single-stream decode

Single-stream decode, shipping engine, same artifact on every platform. Top row is the certified speculative-decoding path, with Neutrino-1 0.6B drafting.

Neutrino-1 8B

Available now

Neutrino-1 0.6B

Available now

Evaluation battery

General knowledge (MMLU)
72.1
Knowledge, re-annotated set (MMLU-Redux)
67.8
Instruction following (IFEval)
77.2
Tool use (BFCL)
68.9

Measured on standard public harnesses, July 2026. Methodology on the model card.

Careers

The frontier of intelligence is now capability per joule, per byte, per second. We hire the people who move it.

Join Fermion Research