The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestFermionResearch2
FermionResearchSpeech → Text

Phonon-2 brings on-device ASR to Apple Silicon

A low-bit quantized, Parakeet-based speech recognizer built to run locally on Mac hardware.

Sep 28, 2026
UpdateOther
Phonon-2

FermionResearch has published Phonon-2, a compact automatic speech recognition model designed to run locally on Apple Silicon. The model is built on NVIDIA's Parakeet ASR architecture and shipped in a low-bit quantized form, with roughly 600 million parameters and a focus on English transcription.

The pitch here is practical rather than flashy: instead of streaming audio to a cloud endpoint, Phonon-2 aims to do the work on the device itself. For developers building Mac-native apps — note-takers, captioning tools, voice interfaces — that means lower latency, no per-request API cost, and audio that never leaves the machine.

Why it matters

Parakeet has earned a reputation as one of the stronger open ASR families, but running it well on consumer hardware has usually meant trimming it down. Phonon-2's low-bit quantization is the lever that makes a sub-gigabyte footprint feasible on Apple's unified-memory chips.

  • Architecture derived from NVIDIA's Parakeet ASR line
  • About 0.6B parameters, quantized to low precision
  • English-only, optimized for on-device Apple Silicon inference

The release is distributed under a non-standard license, so teams planning commercial use should read the terms on the model card before shipping. For anyone who wants fast, private transcription without a server bill, it's a notable addition to the on-device toolkit.

Sources

  • FermionResearch/Phonon-2

    Hugging Face

    Visit
OlderH Company's Holo4 Takes On Computer-Use AgentsH company · Vision-Language · 2 months agoNewerXiaomi Enters Document OCR With New Vision ModelXiaomi · Vision-Language · 2 months ago

Get the model

Hugging Face

Specs

Parameters600M
LanguagesEnglish
Architectureparakeet_tdt_five_value
LicenseOTHER
Downloads3.1K
Likes388

Can you run it?

Runs on just about anything.

  • BF16 (as published)

    Any modern laptop · 8 GB RAM

    2.2 GB
  • FP8

    Any modern laptop · 8 GB RAM

    1.6 GB
Your machine
  • Precision not published; assuming 16-bit (BF16) weights.
  • Parameter count is an editorial estimate — no weights metadata on Hugging Face.

Estimated from the parameter count. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

Speech → Text

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More in Speech → Text

All Speech → Text →
Rumble-2-Flash
Reson8/Speech → TextRuns anywhere

Reson8's Rumble-2-Flash Puts ASR in 0.5B Params

A compact open speech recognition model aims to transcribe more than 40 languages without heavyweight infrastructure.

Nov 18, 2026
Haru-4B
Higurashi Ai/Any-to-AnyRuns on a laptop

Haru-4B Brings Full-Duplex Speech to Six Languages

A compact 4-billion-parameter model handles real-time, two-way voice conversation across six languages in a single system.

Nov 12, 2026
ServiceNow/Text → SpeechConsumer GPU

ServiceNow Debuts Apriel Voice 27B, an Open Speech Model

The 27-billion-parameter preview unifies text-to-speech, transcription, and reasoning in a single open model.

Oct 28, 2026