Laurent Mazare

I have a bunch of catch up to do in ocaml, effects/multicore/oxcaml all seem very exciting to play with!

Thanks for the kind words @anil.recoil.org , just to mention that these models are a team effort from a bunch of folks here at Kyutai! The MLX bugs are most likely mine though :) hopefully I'll get some free cycles soon and make an ocaml implementation that will have less of these!

🚀 Say hello to unmute.sh — a modular voice AI system built on our in-house low latency text-to-speech and speech-to-text engines. It works in English 🇬🇧 and French 🇫🇷 and you can customize the voice and personality.
🎙️Try it live and tell us what you think!

I'll present a dive into Moshi 🟢 and our translation model Hibiki 🇫🇷♻️🇬🇧 as part of the next @convai-rg.bsky.social reading group 👨‍🏫📗.

📅 13th of March 🕰️ 11am ET, 4pm in Paris.

I'll discuss Mimi 🗜️ and multi-stream audio modeling 🔊.
Join on Zoom, replay on YT.

⬛ ⬛ 🟧 🟧 🟨 🟨 🟩 🟩 🟩 ⬛
⬛ 🟧 🟧 🟨 🟨 🟩 🟩 🟩 ⬛ ⬛

Ah, we already have inference codebases in python/rust/swift but maybe we need an ocaml one too for folks that interact with the French ocaml community 😄

Afraid of missing out on French pop culture references because you don't speak the language? Fear no more and try our Hibiki speech-to-speech translation model— no more FOMO! 🇫🇷✨ #Translation #AI

Even Kavinsky 🎧🪩 can't break Hibiki! Just like Moshi, Hibiki is robust to extreme background conditions 💥🔊.

We just released Hibiki 🟢, a real time speech-to-speech translation 🇫🇷 -> 🇬🇧. It preserves the voice of the user, and the smaller variant can run on iPhone as showed by Neil in this video.
Find the code on github github.com/kyutai-labs/... and the weights on HF and give it a spin!

GitHub - nu-dialogue/j-moshi: J-Moshi: A Japanese Full-duplex Spoken Dialogue System J-Moshi: A Japanese Full-duplex Spoken Dialogue System - nu-dialogue/j-moshi

Very impressive to hear this Japanese 🇯🇵 version of moshi 🟢. I don't speak the language so I cannot understand what it's trying to tell me but at least it sounds great 😅
github.com/nu-dialogue/...

Getting our latest LLM to run on the edge was pretty fun, it had been a while since I last used swift and it's still a pretty enjoyable language!

That's certainly the plan, ideally this will make it possible to run a moshi locally on a mobile phone but there is still a lot to be done before we get there!

Super proud of our first publicly released text model, helium-1 preview, a 2B model trained on 6 languages. It should be a great fit for on-device applications. Already available in candle/transformers, can't wait to see what the community builds with it! #OpenSource #AI #FTW!

GitHub - LaurentMazare/ug: Experimental compiler for deep learning models Experimental compiler for deep learning models. Contribute to LaurentMazare/ug development by creating an account on GitHub.

It's certainly open-source but also pretty much a work in progress :) github.com/LaurentMazar... There is no AMD support for now but as long as the programming model is similar to cuda or metal, adding a new backend should be easy (or at least that's the goal).

Ah, sounds like a nice part to have! Note that the bindings are now developed by JS here github.com/janestreet/t... Also lately I've worked on a small compiler for deep learning models, inspired by triton/tinygrad. It's written in rust, but ocaml would be a great choice for writing such a compiler.

The benchmark code with all the matmul variants is on github: github.com/LaurentMazar...

Last week we've received a new M4pro mac mini so I've benched it with various matmul variants and the results are pretty impressive for a tiny form factor. Even with a naive approach it reaches ~5.2TFlops in f32 (so probably more than 10TFlops in bf16), and that's just using the GPU, no NPU for now.