As of: 2026-11-19 22:11 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/m/phonon-2/ # Phonon-2 (open weights, on-device ASR) Fermion Research, Phonon family. Status: current. Type: speech and audio. Released: 2026-09-28. ## How to call it | Provider | Model id | Endpoint | Docs | |---|---|---|---| | Hugging Face | `FermionResearch/Phonon-2` | | https://huggingface.co/FermionResearch/Phonon-2 | | PyPI (CLI) | `fermion-research` | | https://github.com/fermionresearch/phonon https://pypi.org/project/fermion-research/ | | Docker | `ghcr.io/fermionresearch/phonon-cpu:2.0.2` | | | | Mac app (Detta) | | | https://www.fermionresearch.com/products/detta | ## Pricing - Price not published. - Last checked: 2026-09-30 ## Specs - Input: audio - Output: text - Context window: not published - Max output: not published - Open weights: yes - Licence: cc-by-4.0 ## Where it stands - Newer speech and audio models in the registry since its release: 14. The 5 newest: - NVIDIA Nemotron 3.5 Transcribe (30B-A4B) (2026-10-28) https://postcutoff.com/m/nemotron-3-5-transcribe/ - Hy-SimulMT-30B-A3B (2026-10-22) https://postcutoff.com/m/hy-simulmt-30b-a3b/ - Eleven Realtime v1 Max / Mini (2026-10-20) https://postcutoff.com/m/eleven-realtime-v1-max/ - Voxtral Mini Transcribe 3 (+ Mini Transcribe Realtime 3) (2026-10-20) https://postcutoff.com/m/voxtral-mini-transcribe-3/ - Voxtral Small Transcribe 3 (2026-10-20) https://postcutoff.com/m/voxtral-small-transcribe-3/ - Weights are public. Licence: CC BY 4.0. See the licence for commercial terms. ## What stands out - ~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). Source: https://huggingface.co/FermionResearch/Phonon-2 - 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. Source: https://huggingface.co/FermionResearch/Phonon-2 ## 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). Launch post: https://x.com/yoitsmanan/status/2104990913886031993 (≈37k views). Related: [nvidia-parakeet-canary](nvidia-parakeet-canary.md), [whisper-large-v3](whisper-large-v3.md). Model page: https://postcutoff.com/m/phonon-2/