Julian Mack

Happy to share what I've been working on recently: today we release Cohere Transcribe, a state-of-the-art speech recognition model that beats both commercial and open-source models to land at #1 on the Open ASR Leaderboard!

This is cool. Our model running in the browser via transformers.js!

@BarathAnandan7 @huggingface @Tu7uruu @eustachelb @TheOneKloud We did experiment and 35s is the best we saw. It maintains O.K. quality up to around 45s but starts falling off rapidly above that

@Julianfmack @huggingface @Tu7uruu @eustachelb @TheOneKloud Amazing workk :)) Trying to optimize it for mac and spark - is 35s the optimal chunk size or have you experimented with more? Thanks!!

Happy to share what I've been working on recently: today we release Cohere Transcribe, a state-of-the-art speech recognition model that beats both commercial and open-source models to land at #1 on the Open ASR Leaderboard!

Very hyped by the new Cohere Transcribe model 🌍 Works surprisingly well on bad quality audio when the mic doesn't cooperate. 2B params, 14 supported languages and it's Apache 2.0. try the official Hugging Face demo ⬇️

@kushtrimvisoka Our tokenizer does use byte fallback though so while a totally new character set will be challenging for the current vocab, it's not an absolute constraint

@kushtrimvisoka We haven't tested this but we'd be very interested in your results if you try! The main practical constraint is the tokenizer, which covers Latin, Greek, Arabic, Chinese, Japanese kana and Korean Hangul. Adaptation to languages outside these would need tokenizer changes

Cohere Transcribe is the first step of Cohere in the Audio space. Happy to have contributed to this - focused on optimizing the inference stack for speed and scalability.

In addition to the world's best translation models, we make the world's best transcription models at Cohere. This is the best open model out there for English speech recognition, and it works great on other languages too. @Julianfmack and @TheOneKloud absolutely killed it here!

Introducing: Cohere Transcribe – a new state-of-the-art in open source speech recognition.

Excited and proud to introduce our latest: Cohere Transcribe, the best dedicated ASR model in the world. #1 EN HF leaderboard, SotA human evals, ahead of ElevenLabs, Qwen3, Mistral, Kyutai, and OpenAI. 14 supported languages. Apache 2.0, on HF for you to try. Our first audio model and a key step in powering North experiences. huggingface.co/CohereLabs/coh…

Model supports 14 languages and is released open source on hugging face (Apache 2.0). More detail in the technical blog: huggingface.co/blog/CohereLab…

We validated our quality in human preference evaluations. In head-to-head comparisons we come out ahead (>50% win-rate) against all competitors. Meaning preservation was the key criteria. But we also wanted well formatted, verbatim responses with correctly cased proper nouns

@jeankaddour Maybe an annealed aux loss formatting term to put special tokens <start/stop_thinking> in the right place during sft? As the trajectory is ~unchanged after 250, the aux term isn't adding new knowledge vs the baseline

ML interview question: What is happening here?

🚀 Today we’re releasing FlashOptim: better implementations of Adam, SGD, etc, that compute the same updates but save tons of memory. You can use it right now via `pip install flashoptim`. 🚀 arxiv.org/abs/2602.23349 A bunch of cool ideas make this possible: [1/n]

New paper on a long-shot I've been obsessed with for a year: How much are AI reasoning gains confounded by expanding the training corpus 10000x? How much LLM performance is down to "local" generalisation (pattern-matching to hard-to-detect semantically equivalent training data)?

This is the most insane @github thread I’ve ever seen github.com/matplotlib/mat…

Introducing 💡On-Policy Self-Distillation💡, a simple method that enables LLM to teach itself with dense per-token feedback on its own on-policy generations—achieving 4-8x more token efficiency vs. GRPO and outperforming both GRPO and SFT/Off-Policy Distillation. Key insight: like a student reviewing solutions, rationalizing them, and correcting prior mistakes, an LLM can be conditioned on privileged info (e.g., correct solution or a reasoning trace) and supervise its weaker self—the version without such access—by matching the privileged-info-induced distribution from itself. 🌐Blog: siyan-zhao.github.io/blog/2026/opsd/ 🧵👇

Global AI deserves reproducible and transparent evaluation. 🌎 With Global MMLU Lite now part of @kaggle Benchmarks, you can track the multilingual performance of top models as well as test your own! Check out the leaderboard and notebook linked below.

I am hiring highly skilled performance engineers for my team! You will be working on optimising pretraining for models >100B params on O(1000s) of GPUs, and hardware-aligned architecture design. We are cooking a lot of very exciting projects and I can safely say you will have a lot of fun! Link in thread. <3