Phonon-2 (open weights, on-device ASR)
Launch post: https://x.com/yoitsmanan/status/2104990913886031993 (≈37k views). Related: nvidia-parakeet-canary, whisper-large-v3.
Model idFermionResearch/Phonon-2
PricePrice not published
Checked30 September 2026
Where it stands
14newer speech and audio models since its release
- NVIDIA Nemotron 3.5 Transcribe (30B-A4B)28 Oct 2026
- Hy-SimulMT-30B-A3B22 Oct 2026
- Eleven Realtime v1 Max / Mini20 Oct 2026
- Voxtral Mini Transcribe 3 (+ Mini Transcribe Realtime 3)20 Oct 2026
- Voxtral Small Transcribe 320 Oct 2026
The 5 newest are listed. All speech and audio models
- Weights
- Weights are public. Licence: CC BY 4.0. See the licence for commercial terms.
How to call it
4 ways to use Phonon-2 (open weights, on-device ASR), with the model id each provider expects.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face | FermionResearch/Phonon-2 |
huggingface.co | |
| PyPI (CLI) | fermion-research |
github.com pypi.org | |
| Docker | ghcr.io/fermionresearch/phonon-cpu:2.0.2 |
||
| Mac app (Detta) | fermionresearch.com |
Pricing
Price not published.
No price source linked. Checked 30 September 2026.
Specs
- Input
- Audio
- Output
- Text
- Context window
- Not published
- Max output
- Not published
- Open weights
- Yes
- Licence
- CC BY 4.0
- Released
- 28 September 2026
- Status
- Current
- Type
- Speech and audio
What stands out
What its maker marketed at launch, and what people found later, each with its source.
~2-bit quantised Parakeet for on-device English ASR
A quantisation-aware-trained compression of NVIDIA parakeet-tdt-0.6b-v3: encoder weights at one of five learned levels (~2.1 bits), a 164 MB download vs the 2.5 GB teacher, averaging 5.21% WER on the Open ASR Leaderboard’s seven English sets (teacher 4.96%, Whisper large-v3-turbo 6.58%, vendor-run numbers).
Fast local transcription
Vendor figures: about 174x realtime on an M5 MacBook Air (an hour of audio in ~20 s), 143x on eight Zen 5 cores, 6,680x on one H100 at batch 128.
Notes
English only. Not a new architecture: a compressed NVIDIA Parakeet TDT 0.6B v3 (tokenizer and output conventions unchanged). The launch post’s ‘more accurate than Whisper large at 1/10 the size’ refers to whisper-large-v3-turbo in the vendor’s own table; benchmarks not independently reproduced. Fermion Research is a small startup (founder Manan Gupta).