We compress intelligence into a milliwatt.

Fermion trains models natively at extremely low precision, writes its own kernels for each chip, and designs its own weight format. Our ternary format stores each weight in 1.58 bits.

Neutrino-1: an 8B language model in a 2.56 GB download. It outscores Llama-3.1-8B and Gemma-4 on knowledge benchmarks, models that need around 16 GB. (launch, weights)
Phonon-2: a 164 MB speech model that is more accurate than Whisper large and transcribes an hour of audio in 6 seconds on a MacBook Air. It has been the #1 trending speech recognition model on Hugging Face since launch. (launch)
Our runtime: both models run in it on Mac, Linux, Windows and NVIDIA GPUs.
Detta: our Mac dictation app, built on Phonon-2.

This role owns research and engineering across our models: training them, serving them, and building products on top of them.

Research and training

Push low-bit training further: new architectures, formats below 1.58 bits, and keeping accuracy as precision drops.
Train and adapt language, speech and decision models, using pretraining, distillation, preference and RL tuning, and long context.
Build the data pipelines: generation, filtering, decontamination and labelling.
Build the in-house evaluations and hold every release to them.
Run large training jobs on GPU clusters reliably and at sensible cost.

Serving and deployment

Run our models behind our API: inference servers, latency, throughput, autoscaling and GPU cost.
Lead kernel work for each target: CUDA, Metal and MLX, Core ML and the Neural Engine, and CPU.
Deploy custom models for enterprise customers on their cloud, their own servers or their devices, trained on their data.

Products

Work with the product team to build Detta and new consumer apps around our models.
Turn customer problems into models that ship.

About you

You have run inference in production: batching, caching, CUDA graphs, autoscaling, and the cost per million tokens that comes out of it.
You have worked deep in low-bit: ternary or sub-4-bit weights, quantization-aware training, lookup-table formats, or architectures built around them.
You have done serious post-training: SFT, distillation, preference and RL tuning, and you know how to keep a model's capabilities intact through it.
You have shipped quantization that holds up, from GPTQ and AWQ-style post-training methods to QAT and mixed precision, and you know where each one breaks.
You think in rooflines: you know whether a kernel is bound by memory bandwidth or compute on a given chip, what layout and fusion it needs, and how to get it built.
You build evals you would bet a release on: decontaminated, paired, with error bars, and hard to game.
You are comfortable working directly with the product team and with customers.

Bonus: you have trained or deployed speech recognition models (CTC, RNN-T or TDT), or decision models that return calibrated choices, scores and classifications instead of free text. Experience with distributed training at scale also helps.

What we offer

Founding equity and a competitive salary.
Real compute, and the freedom to use it.
Top-end hardware and tools for whatever the role requires.

Apply

Email careers@fermionresearch.com with a link to something you have built, a few lines on the hardest model problem you have solved, and your CV or GitHub.

Founding Researcher & Engineer