Docs
Browse documentation

Docs / Getting started

Install and run Neutrino

Install fermion, open a local chat, and verify the downloaded model in about a minute.

Requirements

Fermion supports Python 3.10 through 3.13 on macOS, Linux, and Windows. The package is Apache-2.0 licensed and installs the fermion command. On Linux x86-64 the CPU speech engine needs glibc 2.34 or newer (Ubuntu 22.04, Debian 12, RHEL 9 and later) and about 7 GB of free RAM.

shell
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install fermion-research

Install and chat

shell
pip install fermion-research
fermion chat

The first command installs the package. The second downloads Neutrino-8B when needed, verifies it, selects the best available backend, and opens the interactive prompt.

  • /exit leaves the chat.
  • /reset clears the conversation history.
  • The default model is FermionResearch/Neutrino-8B.

Generate one response

shell
fermion generate "Explain why the sky is blue." --max-new 64

generate is deterministic by default and prints the completion to standard output. Runtime and throughput statistics go to standard error, which keeps scripts clean.

Verify the installation

shell
fermion --version
fermion info

fermion info checks the container structure, expected length, and SHA-256. It exits with status 1 if the container is truncated or altered.

What happens on first run

  1. 1

    Fetch metadata

    Configuration, tokenizer files, the model manifest, license, and the runner for the current platform arrive first.
  2. 2

    Download weights

    Apple Silicon downloads a 2.56 GB coded transport and expands it locally. Other platforms download the 3.88 GB container.
  3. 3

    Verify

    The expanded container is checked against the released SHA-256 before it is loaded.
  4. 4

    Select a backend

    The CLI reports whether it selected the optimized native path or the PyTorch reference path.

Speech

shell
fermion transcribe recording.wav
fermion listen

The same package runs Phonon-1, an open speech recognition model for English. transcribe prints the transcript of an audio file to standard output. listen transcribes the microphone live and prints the full transcript when you stop with Ctrl-C. Phonon-1 downloads on first use in 415 MB and runs on a Mac (via MLX), an NVIDIA GPU, or a CPU.

Transcribe and dictate covers both commands, the audio API, and the model options.