Docs
Browse documentation

Docs / Reference

CLI reference

Every command and the shared runtime, sampling, context, and backend flags in fermion 0.1.24.

Commands

CommandPurposeCommand-specific options
fermion chatInteractive streaming REPL.--draft
fermion generateOne-shot completion, greedy by default.prompt, --raw, --draft
fermion transcribeOne-shot speech-to-text for an audio file.audio, --model, --json, --verbose, --download-only
fermion listenLive microphone transcription.--model, --wav, --verbose
fermion infoHeader, completeness, SHA-256, and selected backend.--no-checksum
fermion serveOpenAI-compatible local HTTP service.host, port, auth, sessions, tools, draft
fermion inspectContainer geometry, occupancy, bytes, and KV cost.--layer, --projection, --sha256, --json
fermion verifyCompare plain and drafted token streams.--draft required, --prompts
fermion benchSingle-stream local benchmark.--tokens, --prompt, --warmup, --draft

Shared flags and defaults

FlagDefaultMeaning
--modelFermionResearch/Neutrino-8BA local artifact, model directory, or Hugging Face repository id.
--devicecpucpu, mps, or cuda. Native execution requires cpu.
--dtypeautoPyTorch activation dtype: auto, bfloat16, float16, or float32.
--max-new256 generate; 1024 chat/serveMaximum generated tokens.
--temperature0 generate; 0.01 chat/serveZero selects deterministic argmax.
--top-p1.0Nucleus sampling when temperature is above zero.
--rep-penalty1.0 generate; 1.05 chat/serveMultiplicative repetition penalty.
--pen-window256Recent-token window used by repetition penalty.
--backendautoauto, native, or torch. native fails if the optimized runner is unavailable.
--native-binnoneUse a specific native runner binary.
--kv-dtypefp16Native KV storage: fp16, fp32, or int8.
--yarn-factoroffExperimental native-only RoPE scaling.
--yarn-orig-maxmodel maximumOriginal context boundary used by YaRN.

Serve-only flags

FlagDefaultMeaning
--host127.0.0.1Bind address.
--port8000Listening port.
--draftnoneDraft path, model id, or auto.
--api-keynoneRequire a bearer key on /v1 routes.
--corsoffAllow cross-origin browser requests.
--served-model-namecontainer filenameModel id reported by the API.
--max-new-ceiling4096Clamp client token requests to a server maximum.
--session-ctx8192Resident native KV capacity.
--no-sessionoffDisable the resident session runner.
--stuck-detectorenforceenforce, observe, or off.
--stuck-repeat3Repeated identical-call threshold.
--stuck-alt6Strict alternation window.
--stuck-error2Repeated identical-error threshold.
--stuck-whitelistemptyComma-separated tools that may repeat.
--stuck-no-auto-whitelistoffDisable automatic poll-tool detection.
--tool-profilegradedgraded, vendor, vendor-thinking, or antiloop.

Speech commands

shell
fermion transcribe clip.wav
fermion listen --wav clip.wav
fermion serve --model phonon
FlagDefaultMeaning
--modelFermionResearch/Phonon-1Speech repository id, alias, or a local model directory.
--jsonofftranscribe: emit a JSON object with the text, model id, and decode time.
--download-onlyofftranscribe: fetch and verify the model, print its directory, and stop.
--wavnonelisten: stream a file through the live path, paced to real time, instead of the microphone.
--verboseoffPrint the decode configuration and timings to standard error.

Both commands decode greedily with no sampler flags, and standard output carries only the transcript; progress and warnings go to standard error. Exit status is 0 on success, 1 for a refusal or failure with one plain message on standard error, and 2 for a usage error.

With a speech model, fermion serve mounts POST /v1/audio/transcriptions and GET /v1/audio/stream instead of the chat routes. --host, --port, --api-key, --cors, and --served-model-name keep their meaning; the sampler and backend flags do not apply, and the server says so once at startup and ignores them.

Built-in help

shell
fermion --version
fermion --help
fermion serve --help