Building @fermion_ai, also i am very cool. @mercatus fellow. @solofounders cohort 5 , prev @fdotinc @localhosthq
TON-618
Joined May 2025
- Tweets367
- Following269
- Followers2.9K
- Likes816
Pinned Tweet
Introducing Phonon-2: a new standard in speech recognition per byte.
At just a tiny 164MB download, it is more accurate on average than OpenAI's Whisper large ( a model 10X its size)
Transcribe an hour of audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Manan Gupta retweeted
Bark 3 is English-only, so it doesn't belong on a 25-language chart. And on English, Phonon-2.5 English beats it anyway, 4.82% vs 4.94% average WER on the Open ASR Leaderboard's seven sets.
Also worth knowing that Bark 3 is built on the five-value weight format we introduced with Phonon-2. Same scheme, same tensor layout, and their Hugging Face model card copies ours almost word for word with zero credit given
Introducing Phonon-2.5: A new accuracy-to-size Pareto frontier in multilingual speech recognition.
It is a miniscule 156MB download, which is more accurate on average, than OpenAI’s Whisper large-v3 ( 20x its size!)
Transcribe an hour of multilingual audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Meet the Phonon-2.5 Family, explore the models here:
fermionresearch.com/research…
Related Links:
Announcement: fermionresearch.com/research…
Model: huggingface.co/FermionResear…
Model (English): huggingface.co/FermionResear…
HF: huggingface.co/fermionresear…
pip: pypi.org/project/fermion-res… — pip install fermion-research
Docker (CPU): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
Docker (NVIDIA): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
GitHub: github.com/fermionresearch/p…
Detta, dictation for the Mac: fermionresearch.com/products…
Site: fermionresearch.com
Join us: fermionresearch.com/careers/
Manan Gupta retweeted
Phonon-2 downloads have grown 22X over the past week across HF, our CLI, and community ports.
It's also already in use in 100+ open source projects with over 40k+ combined stars, and deployed in 50+ enterprises.
Our customers have seen a 90%+ reduction in inference bills.
We'll work closely with your team to deploy Phonon, our other models, or specialized models trained on your production traces, at the lowest latency and cost per request of any deployment option.
Contact us here!
fermionresearch.com/deploy/
We worked extremely closely with the Apple Neural Engine to make Detta ridiculously fast, a typical dictation length can be transcribed in a median 11 ms with polish off!
We've also improved Detta's hotword biasing for names and specialized terminology. On our hotword test, custom-term recall jumped from 54% to 79%, with zero false substitutions, and there's a lot more coming here in the next updates.
Read our in depth dive into Detta here.
fermionresearch.com/research…
We are hiring across Engineering and Research teams, if working on the most awesome of problems excites you, shoot us an email at [email protected] with your coolest work.
Related Links:
CoreML Phonon-2: huggingface.co/FermionResear…
Detta: fermionresearch.com/products…
CoreML Phonon-2 Swift Package:
github.com/fermionresearch/p…
pip: pypi.org/project/fermion-res… — pip install fermion-research
Docker (CPU): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
Docker (NVIDIA): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
GitHub: github.com/fermionresearch/p…
Research: one eighth the bits: fermionresearch.com/research…
Site: fermionresearch.com
Join us: fermionresearch.com/careers/
We’re releasing the Core ML package for Phonon-2, reaching a staggering 606× real time, up from our 174× launch figure.
That’s an hour of speech in just 6(!) seconds, on a base M5 MacBook Air, with no loss in quality.
It’s now the default in Detta. We’ve cut peak app memory from 3.1 GB to 1.1 GB, and idle memory from 1.8 GB to just 600 MB, even with Gluon loaded!
That puts Detta's peak memory footprint same as WisprFlow, with the models in Detta running completely locally!
Gluon, our text polishing model, turns messy dictation into clean, formatted text.
Say “Let’s meet on Tuesday, actually no, Wednesday at noon,” and get “Let’s meet onz Wednesday at noon.” It removes the “ums,” repetitions and false starts, adds punctuation, and formats numbers and dates.
Download and update Detta for free today.
fermionresearch.com/products…
Related Links:
Announcement: fermionresearch.com/research…
Model: huggingface.co/FermionResear…
HF: huggingface.co/fermionresear…
pip: pypi.org/project/fermion-res… — pip install fermion-research
Docker (CPU): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
Docker (NVIDIA): ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
GitHub: github.com/fermionresearch/p…
Detta, dictation for the Mac: fermionresearch.com/products…
Research: one eighth the bits: fermionresearch.com/research…
Site: fermionresearch.com
Join us: fermionresearch.com/careers/
Introducing Phonon-1: a new frontier in speech recognition per byte.
A 782 million param model in a 415 MB download, more accurate than a Whisper 2x its size.
Transcribe an hour of audio in about two minutes on a MacBook Air.
Open weights, Apache 2.0, today.
@fermion_ai
Meet Phonon-1. Explore the model and run it today.
fermionresearch.com/research…
Related links:
Announcement: fermionresearch.com/research…
Models: huggingface.co/fermionresear…
pip: pypi.org/project/fermion-res… — pip install fermion-research
Docker: ghcr.io/fermionresearch/phon… — docker pull ghcr.io/fermionresearch/phon…
GitHub: github.com/fermionresearch/p…
Research: one eighth the bits: fermionresearch.com/research…
Site: fermionresearch.com
Join us: fermionresearch.com/careers/