Docs / Getting started

Models, downloads, and storage

Know what the CLI downloads, where it stores the files, and how much free space the first run needs.

Released models

Model	Runnable container	Coded transport	Primary use
Neutrino-1 8B	3.88 GB	2.56 GB	Default general model and tool caller
Neutrino-1 0.6B	328 MB	238 MB	Small model and canonical speculative draft
Neutrino-1 0.6B-Chat	328 MB	238 MB	Small conversational model

Speech models

The speech commands default to FermionResearch/Phonon-1. Every speech command takes --model with a repository id, a short alias, or a local model directory.

Model	Download	On disk	Aliases
Phonon-1	415 MB	455 MB	phonon, phonon-1, stt, asr
Phonon-1-Big	581 MB	822 MB	phonon-1-big, big
Phonon-1-Micro	285 MB	331 MB	phonon-1-micro, micro

shell

fermion transcribe clip.wav --model micro
fermion serve --model FermionResearch/Phonon-1-Micro

A speech model downloads as a single verified archive and unpacks under ~/.cache/fermion/speech; the archive itself sits in the Hugging Face hub cache. FERMION_CACHE_DIR and HF_HOME move both.

The first-run download plan

The CLI downloads only the files required to run the selected model. GGUF and MLX files are not fetched by default.

Platform	Downloaded	Steady state	Peak free space
Apple Silicon	2.56 GB transport	about 3.9 GB	about 6.5 GB
Other platforms	3.88 GB container	about 3.9 GB	about 4 GB
0.6B models	238 MB transport or 328 MB container	328 MB	under 600 MB

On Apple Silicon, the verified transport is deleted after expansion. Set FERMION_KEEP_TRANSPORT=1 if you want to keep it. Set FERMION_DOWNLOAD_ALL=1 only when you intentionally want the whole repository.

Move the model cache

Models use the standard Hugging Face cache under ~/.cache/huggingface/hub. Move all Hugging Face data with HF_HOME, or move only Fermion model downloads with FERMION_CACHE_DIR.

shell

export HF_HOME=/big/disk/hf
# or
export FERMION_CACHE_DIR=/big/disk/fermion

fermion chat

A small ~/.cache/fermion directory may also hold generated Transformers configuration directories.

Disk-space checks and interrupted downloads

Before downloading weights, fermion checks the real cache volume against the model manifest. If the volume is too small, it stops before the transfer and prints the cache path, required bytes, available bytes, and the two cache variables above.

If a download or local expansion is interrupted, run fermion info. The command detects incomplete structure and hash mismatches and exits non-zero for automation.