---
title: "Changelog"
description: "Recent changes to the public fermion package, verified against the shipped package source."
canonical: "https://www.fermionresearch.com/docs/changelog/"
source: "Fermion Research"
---

Docs / Reference

# Changelog

Recent changes to the public fermion package, verified against the shipped package source.

## 0.2.15

- `fermion serve` on an address other than localhost now requires `--api-key`.
- Phonon models load faster on Apple silicon; long recordings use less memory.
- Containers: `phonon-cpu:2.0.12` (amd64 and arm64) and `phonon-cuda:1.0.11` run this release. `phonon-cuda` starts much faster and transcribes longer files.
- Performance improvements and bug fixes.

## 0.2.14

- Fixes a crash when a model is used from two threads at once, or loaded again in the same process.
- `fermion serve` accepts long recordings as uploads.
- Containers: `phonon-cpu:2.0.11` (amd64 and arm64) and `phonon-cuda:1.0.10` run this release.
- Performance improvements and bug fixes.

## 0.2.13

- Phonon-2.5, speech recognition in 25 languages: `fermion transcribe phonon-2.5 recording.wav`.
- Phonon-2.5 English: `fermion transcribe phonon-2.5-en recording.wav`.
- Smaller model downloads for the Phonon models; the same files on disk after unpacking.
- Performance improvements and bug fixes.

## 0.2.12

- Neutrino models load through `transformers` exactly as the model cards show, with `import fermion` and then `from_pretrained`.
- They compute in bfloat16 by default on every device, NVIDIA GPUs included.
- The Neutrino-8B runtime for Linux on x86-64 takes `--exact`, for greedy output that matches the float32 reference token for token.
- Performance improvements and bug fixes.

## 0.2.11

- Performance improvements and bug fixes.
- Containers: `phonon-cpu:2.0.9` (amd64 and arm64) and `phonon-cuda:1.0.8` run this release.

## 0.2.10

- Faster on Intel and AMD processors. Phonon-2 loads much faster and decodes faster on Linux and on Intel Macs, and uses less memory while it runs. Transcripts are unchanged, byte for byte.
- Segment timestamps. `segments` holds one entry per sentence or pause-sized stretch of speech, with `start` and `end` in seconds from the start of the file and its text; the segment texts joined with single spaces give the transcript. `fermion transcribe phonon-2 meeting.wav --json` returns them, and `fermion serve` fills `segments` in `verbose_json` in the OpenAI shape.
- One-line install. `pip install fermion-research` brings everything the CPU engine needs on Linux, Windows and Intel Macs. On Apple silicon, `pip install "fermion-research[mlx]"` adds the MLX speech runtime.
- `FERMION_CPU_TIER` runs the tier it names.
- Containers: `phonon-cpu:2.0.8` (amd64 and arm64) runs this release; in `phonon-cuda:1.0.7`, `transcribe --json` returns segments and word timestamps.

## 0.2.9

- Hotwords on every Phonon-2 engine. Give Phonon-2 up to 25 names and terms (people, products, places, jargon) and it favours them whenever the audio is close; nothing is ever penalised, so the audio still decides. Without hotwords every transcript is unchanged.
- `fermion transcribe` and `fermion listen` take `--hotwords` (or `--hotwords @names.txt`); the Python API takes `hotwords=[...]`; `--hotword-lambda` (`hotword_lambda=`) sets how hard the listed words pull.
- The server honours the OpenAI `prompt` field as a vocabulary list and takes an explicit `hotwords` field. Each request’s words apply to that request only.
- Hotwords work on Apple silicon, on Linux, Windows and macOS CPUs, and in the CUDA image `phonon-cuda:1.0.6`.

## 0.2.8

- Intel Macs run Phonon-2 through the CPU engine. The engine is chosen on its own; `fermion describe` shows the tier (AVX2 on Core i5/i7/i9 and Xeon Macs, a baseline tier on parts without it).
- Python 3.10 to 3.12 on Intel Macs; the package resolves `torch<2.3`, `numpy<2` and `transformers<5.1` on that platform only.
- The Phonon-1 family is not available on Intel Macs; the command says so in one line, before any download.
- First published on PyPI as part of 0.2.9.

## 0.2.7

- Word timestamps. `fermion transcribe phonon-2 clip.wav --json` returns `words`, one entry per word with its start and end in seconds from the start of the file, placed correctly on long audio.
- `fermion serve` honours `timestamp_granularities`: with `word`, the `verbose_json` response carries `words` and `segments`.
- Both Phonon-2 engines return timings, the Apple-silicon engine and the CPU engine on Linux, Windows and macOS. Transcripts are unchanged.

## 0.2.6

- `fermion serve` with a name that is not a published model or a local directory stops in one line before any network request.
- `fermion transcribe` checks that the audio file exists before it loads a model.
- `fermion models` names the model in its example commands; help texts describe the package, the CPU platforms and the model-first `serve` form.

## 0.2.5

- The CPU engine reads the processor’s features before it loads a kernel (dotprod, i8mm and SVE on 64-bit Arm; SSE4.1, AVX2, AVX-512 and AMX on x86-64) and picks the matching tier.
- Arm processors without dotprod (Raspberry Pi 3 and 4, Cortex-A53 and A72 boards) and x86-64 processors without AVX2 run Phonon-2 on a baseline tier instead of stopping.
- `fermion describe` and `phonon describe` print the processor features found, the tier chosen and the binaries loaded, with `--json` for scripts.
- The package carries its Apache-2.0 licence text and NOTICE file.

## 0.2.4

- Every command names its model, as in `phonon transcribe meeting.wav` or `fermion transcribe phonon-2 meeting.wav`. A command without a model prints the model names and exits.
- The `phonon`, `phonon-2` and `phonon-1` commands run a fixed model.
- `fermion serve` takes `--unix-socket` for an owner-only local socket and `--threads` for the CPU engine.
- The containers phonon-cpu 2.0.3 and phonon-cuda 1.0.4 take the model first and keep it in the `phonon-cache` volume.

## 0.2.3

- The command line checks it can read audio before downloading a model.
- `fermion models` describes each model in plain words.
- The README lists the CPU and NVIDIA container images.

## 0.2.2

- `fermion serve` with its default model starts again with transformers 5.17.
- Windows: `fermion speech bench` prints on consoles that cannot show Unicode.
- Linux: the glibc note appears before any download.

## 0.2.1

- Windows: `fermion` commands start again; in 0.2.0 every command stopped as soon as it was launched.
- Linux: on glibc older than 2.28 the CPU engine says so and, once libsndfile1 is installed, runs a slower full-precision path instead of failing to load.
- Documentation: the README names Phonon-2 as the default speech model, lists the names that select each model (`phonon-2`, `phonon2`, `phonon`, `speech`, `stt` and `asr` for Phonon-2; `phonon-1` for Phonon-1), uses `phonon-2` in the serve example, and describes the CPU speed machine as eight cores (16 vCPU).
- Known: `fermion serve` with its default model does not start on macOS and Linux x86-64 with transformers 5.17; use `fermion serve phonon-2`, or pin `transformers<5.17`.

## 0.2.0

- New speech model: Phonon-2 (`FermionResearch/Phonon-2`), selected with `--model phonon-2`. It is a 164 MB download, and it averages 5.21 % word error on the seven public test sets of the Open ASR Leaderboard. On an M5 MacBook Air it transcribes a single stream at 174 times realtime. The weights are released under CC-BY-4.0.
- Phonon-2 is now the default speech model. `fermion transcribe` and `fermion listen` without `--model` run it, `fermion serve phonon-2` serves it, and the name `phonon` selects it everywhere. Phonon-1 remains available as `--model phonon-1`.
- Phonon-2 runs on CPUs on macOS, Linux (x86-64 and ARM) and Windows. On eight Zen 5 cores (16 vCPU) it transcribes a single stream at 143 times realtime, at the accuracy of full precision (3.94 % word error on LibriSpeech test-other against 3.91 %); 52 times on Google Axion, 40 times on the CPU of an M5 MacBook Air. `FERMION_DEVICE=cpu` selects the CPU engine on a Mac.
- Long recordings are cut at pauses, and each window is decoded with 0.25 s of silence at both ends so that words at the edges stay whole.
- `fermion speech bench --audio recording.wav` reports load time, milliseconds of compute per second of audio, the realtime factor and peak memory for the selected model on the current machine.
- Requirements: Python 3.10 to 3.14. Apple silicon runs the MLX engine; the CPU engine on Linux needs glibc 2.28 or newer (Ubuntu 18.10 or newer, Debian 10+, RHEL 8+).

## 0.1.24

- Fixes a startup failure on systems with a broken torchaudio install; the runtime now starts and tells you how to remove the broken package.
- Requirements: 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. macOS and Windows requirements are unchanged.

## 0.1.23

- Long recordings are transcribed in full. `fermion transcribe` and `POST /v1/audio/transcriptions` decoded a whole file in one pass with a token budget sized for a single utterance, so anything over about two minutes came back truncated and anything over about three degenerated into a repeated letter. Audio longer than 35 s is now cut into 25 to 35 s windows at pauses and each window is decoded on its own, on both the MLX and the CPU engine. Audio up to 35 s takes the unchanged single-pass path, so short-utterance transcripts are byte-identical to 0.1.22.
- `--json` adds `wall_seconds` next to the decode-only `decode_seconds`, per-segment start and end times, `segment_count`, and `truncated`. `--verbose` prints the same, labelled. `verbose_json` on the server fills `segments` and reports `truncated`.
- One always-on line at engine start names the decode backend (mlx or cpu) and the model, and says on the first run in a new environment that the run compiles once and later runs are fast.
- A Python running under Rosetta on an Apple-silicon Mac is told so, with the fix (an arm64 Python), instead of being refused as an Intel Mac.

## 0.1.22

- The CPU speech engine now also runs on ARM Linux (aarch64): `fermion transcribe` and a speech `fermion serve` work on 64-bit ARM Linux machines with the same install line as x86-64 Linux. The models are the same verified downloads, and transcription stays deterministic.
- The CPU engine's first run no longer prints a spurious advisory line from the model backend's import.
- `fermion models`: one transcribe example in the usage footer.

## 0.1.21

- Speech on plain CPU: `fermion transcribe` and a speech `fermion serve` now run on x86-64 Linux and Windows, and on a Mac without the MLX runtime. `pip install fermion-research torch safetensors soundfile scipy zstandard` is the whole setup; the models are the same verified downloads. Apple silicon keeps the MLX engine as its default, and `FERMION_DEVICE=cpu` selects the CPU engine explicitly.
- Transcription decodes deterministically on every engine; transcripts across engines are same-accuracy, not word-identical.
- `fermion models` reports which speech engine this machine runs.

## 0.1.20

- Package documentation copy edits; no behavior change.
- `fermion models`: shorter speech model summaries.

## 0.1.19

- Package documentation refreshed for the Phonon-1 release.

## 0.1.18

- Streaming speech over WebSocket: a speech `fermion serve` also mounts `GET /v1/audio/stream`, taking one JSON config frame and then raw mono audio, and answering with partial hypotheses, final segments, and the full transcript on close. `--api-key` applies, as a Bearer header or `?api_key=` for browser clients.
- `fermion serve --model` pointed at a local speech model directory now serves speech; it previously failed down the chat path.

## 0.1.17

- `fermion listen`: live dictation from the microphone. Partial hypotheses update one terminal line, a pause finalizes the segment, and Ctrl-C prints the full transcript to standard output.
- `fermion listen phonon-2 --wav FILE` streams a file through the same path paced to real time.
- `fermion --version` reads the installed package metadata instead of a maintained literal.

## 0.1.16

- Fresh-install fix: `mlx-lm` added to the checked speech runtime and the documented install line. Without it, a first run crashed after the model download instead of printing the install hint.

## 0.1.15

- Speech support: `fermion transcribe` and speech serving with `/v1/audio/transcriptions`.
- Three speech models, downloaded on first use and selected with `--model`: `FermionResearch/Phonon-1` (the default), `Phonon-1-Big`, and `Phonon-1-Micro`.

## 0.1.14

- Native `--draft` routing with no PyTorch model loaded on the optimized path.
- `--draft auto` resolves to Neutrino-0.6B.
- `fermion verify` defaults to float32 so PASS matches the identity guarantee.
- Clearer reduced-precision near-tie messages.

## 0.1.13

- More tolerant multi-call envelope parser and private think-block guard.
- Serve-side stuck detector and tool-request sampling profiles.
- Warnings for unsupported OpenAI sampling fields.
- TF32 disabled during CUDA float32 verification.

## 0.1.12

- Int8 KV column in `inspect`.
- Clearer YaRN scope and per-model tool-calling guidance.

## 0.1.11 and earlier

- 0.1.11 added free-space preflight, named disk errors, `FERMION_CACHE_DIR`, and post-expansion transport cleanup.
- 0.1.10 added KV dtypes, experimental YaRN controls, and session memory guards.
- 0.1.9 added coded Apple Silicon downloads and live prose streaming for tool requests.
- 0.1.8 added unique tool-name aliasing, multiple KV conversations, and session statistics.
- 0.1.7 added OpenAI tools, persistent KV sessions, context clamps, and longer chat output.

Source: [https://www.fermionresearch.com/docs/changelog/](https://www.fermionresearch.com/docs/changelog/)
