AI Research engineer โ Building apps like Scribe Instant Transcrips, VoiceView, ScreenRun โ MTB ๐ต ๐๏ธ
Palo Alto and Chamonix
Joined April 2007
- Tweets7K
- Following2.5K
- Followers1.2K
- Likes17.5K
test Kokoro, Supertonic, PocketTTS, Phonon and Pico Voice TTS models running 100% in your browser at ldenoue.github.io/xn-ptts/
Laurent Denoue retweeted
My newest book 'Fine Tuning AI' just came through the mail from O'Reilly
I think this is perhaps the most important book I've ever written. I strongly believe the best future career for many software developers will be in Fine Tuning AI for domain-specific tasks. With this book, I try to simplify it, primarily using Google's Gemma models.
You can find it on Amazon here: link.amazon/B02x2ZVPD
Or O'Reilly Safari here: oreilly.com/library/view/finโฆ
Thank you so much for your support!
NiceGrab for macOS: 4 promo codes:
appblit.com/nicegrab
PARYEX6F9JX4
3ARA64F3P6Y6
7K767H9JNHXL
FHH6T493TPXH
PopMath Original for iOS and iPad was just updated. This app won many awards: yet it took 4 hours to code back in 2008 when the App Store first opened. apps.apple.com/us/app/popmatโฆ
New version of PopGeo USA Geography for 9-12 kids to learn state capitals, cities, national parks, all beautifully illustrated with stamps presenting each. appblit.com/popgeo
Laurent Denoue retweeted
A quarter of a second. That's how long four of our on-device models took to turn a 12-second voice memo into a named, sorted note.
Voz transcribed the audio, Title named it, Emo picked an icon and Gist filed it under a topic.
Chain them together in your app and you have a voice notes feature where the recording never leaves the device.
Laurent Denoue retweeted
T-minus 6.4 seconds.
That's how long our speaker diarization model, Who, took to label every speaker when we tested it on a 57-minute NASA podcast.
It ran at 530ร real time, on a MacBook, with nothing leaving the device.
Laurent Denoue retweeted
It's insane that you can now run your own coding agent (Pi) w/ LFM3-1.7B *in the browser* at over 300 tok/s on an M4 Max, using custom WebGPU kernels & 4-bit GGUF weights.
No data sent to a server. All your chats stay private.
Try it yourself on Hugging Face ๐
Laurent Denoue retweeted
Introducing ๐ป๐ต๐ผ๐น: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding
> ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language.
> ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ ๐งต
Incredible how optimizing image and video pre-processing to run RVM (background segmentation) LiteRT models can make your FSP jump from 6 (WASM) to 30 (LiteRT WebGPU) to 60 FPS (LiteRT WebGPU + optimized GPU video textures to avoid CPU roundtrips) ldenoue.github.io/rvm-webgpuโฆ
Nicegrab for macOS (clean screenshots and recordings with a simple key shortcut, composer over a background and smooth cursors) macmenubar.com/nicegrab/
100% browser based immersion of several video sources: recoloration based on ref image, relighting, segmentation, shadows, light sources and music beat detection driving the light intensity.
Now we just need to make it run at 60 FPS (possible on iOS) and of course RVM would produce much nicer mattes than the coarse masks given by MediaPipe (but RVM is slower!!)
Laurent Denoue retweeted
Today I'm releasing Paradee-8M-v1.1, a new version of my small open-weight text-to-speech model, distilled from Kokoro-82M. It's 8.5MB and it runs on any laptop without a GPU.
In v1.1, Paradee now runs 62x faster than real time on one CPU thread (compared to 18x from v1.0), and provides cleaner audio.
Laurent Denoue retweeted
Ray Dalio on the advice he gives his grandkids.
@RayDalio
Laurent Denoue retweeted
US: Tesla Roadster. EU: tethered bottle caps.
I made that joke too. But let me be serious for a moment.
Made in Europe.
Europe built it.
Others told the story.
The Web was born at CERN.
Linux in Helsinki.
Python in Amsterdam.
ARM, the chip in your phone and your Mac, in Cambridge.
DeepMind in London.
The first commercial microprocessor was designed by an Italian, Federico Faggin.
Every leading-edge chip on Earth is printed with ASML machines from the Netherlands.
Mistral trains frontier open models in Paris.
Hugging Face: French founders, although it is headquartered in New York.
llama.cpp, the engine of local AI, started in Bulgaria.
The first mRNA vaccine came from Mainz.
Europe doesnโt lack innovation.
It lacks a megaphone. ๐คท๐ปโโ๏ธ
Innovation without narrative becomes someone elseโs success.
This isnโt Europe vs. anyone.
I love innovation wherever it happens.
I just want Europeans to feel proud and optimistic about the future again.
Did I miss something relevant here?
