🤷 Nico Martin

Building Transformers.js at 🤗 @huggingface // @GoogleDevExpert for #WebTechnologies and #AI

[🤷 Nico Martin @nicodotdev]

@maiieul Thats actually from our balcony😍

[🤷 Nico Martin @nicodotdev]

I'm sooooo soo happy to see the first snow in the mountains😍

[🤷 Nico Martin @nicodotdev]

Your turn: can you get 100/100?

Try heads or tails, then build a tiny JSON creature one token at a time.

@googlegemma 4 E2B runs locally in your browser via Transformers.js + WebGPU.

Play Token Arcade:
https://t.co/arExB56U5b

[🤷 Nico Martin @nicodotdev]

I turned structured output into a game 🎮

The LLM scores the next tokens. You pick one. The constraints decide if it’s allowed.

Most likely ≠ valid. Even a capital letter can lose you points.

Watch the demo. Play it yourself below 👇

[🤷 Nico Martin @nicodotdev]

New York, here I come! 🗽

On Nov 16 I'm speaking at JSNation US about multimodal AI agents that see, hear and remember, all running entirely in your browser.

[JSNation | The Main WebDev Conference @thejsnation]

Voice, vision and memory, all running in the browser. No backend, no API keys 🙌

@nicodotdev, Machine Learning Engineer at @huggingface, shows how WebGPU and WebAssembly make real-time AI agents possible on the open web.

Join us at JSNation US 👇
https://t.co/rpeAm4rV4u

[🤷 Nico Martin @nicodotdev]

For all the german speaking people, I'll be on the token therapie podcast with @janschultecom and Michael Köpf live in about 5 minutes: https://t.co/kjvLK3p306

[🤷 Nico Martin @nicodotdev]

More Ternary Bonsai 2 models from @PrismML just dropped: 2B, 4B, 9B, 27B 🌱

Hundreds of tok/s on consumer hardware, fully in the browser with custom WebGPU kernels. At ~1.8 bits/weight the 2B is just 430MB. Local agentic AI is getting wild 🤯

[Xenova @xenovacom]

Today has been one of the biggest days for open-source AI this year... and it keeps getting better.

@PrismML just released more Ternary Bonsai 2 models on Hugging Face: now in 2B, 4B, 9B, and 27B. The 2B reaches up to 500 tok/s on my M4 Max ENTIRELY in the browser with custom WebGPU kernels... that's insane! 🤯

At ~1.8 bits per weight, the language-model weights are just 430MB for 2B and 950MB for 4B (plus ~365MB for the 8-bit vision weights). The future of local agentic AI is going to be wild.

Try out the demo yourself 👇

[🤷 Nico Martin @nicodotdev]

I'll bring some input on running AI directly in the browser, but mostly I’m looking forward to the conversations.

Got a topic? Submit a session. Entry is free.
https://t.co/JLq3ZBEbxE

[🤷 Nico Martin @nicodotdev]

Next week I'll be at F{AI}R, an AI unconference in Liebefeld, Bern 🇨🇭

The best part of unconferences: the schedule is built together on the spot, and the best moments usually happen over coffee, not on stage.

[🤷 Nico Martin @nicodotdev]

@joelniklaus Huge congrats on the outstanding paper award, Joël! 🎉 Really well deserved, and so good to see data work getting this kind of recognition 🤗

[🤷 Nico Martin @nicodotdev]

The prompt says what we want. The constraints define what’s allowed.

Token Arcade lets you step through that distinction yourself: probabilities, token choices, rejected continuations, and explanations.

Code:
https://t.co/xXoiaWpVeK

[🤷 Nico Martin @nicodotdev]

Your turn: can you get 100/100?

Try heads or tails, then build a tiny JSON creature one token at a time.

@googlegemma 4 runs locally in your browser via Transformers.js + WebGPU. First visit downloads the model.

Play Token Arcade:
https://t.co/arExB57rUJ

[🤷 Nico Martin @nicodotdev]

@KarnikShreyas That is amazing 🚀!

[🤷 Nico Martin @nicodotdev]

Multimodal similarity search in the browser❤️!

[Xenova @xenovacom]

NEW: EmbeddingGemma 2 just released on Hugging Face. Project text, images, video, and audio into a single, unified embedding space.

At just 740M parameters, it can even run 100% locally in your browser on WebGPU with 🤗 Transformers.js!

Try it yourself 👇

[dotConferences @dotConferences]

AI that runs in your browser, on your hardware, without calling the cloud? That’s not boring. That’s pretty handy.

In “Boring but Effective: WebAI That Ships,” Nico Martin shows how small open models can handle useful tasks right in the browser - from semantic search to speech recognition and background removal. No API key, no cloud subscription, no giant download. Just WebAI that’s ready to get to work.

@nicodotdev is a Machine Learning Engineer focused on WebML at Hugging Face, and a Google Developer Expert in AI and Web Technologies. He makes it easier for web developers to bring machine learning into the browser - and shares what he learns through workshops, articles and talks.

🟡 Catch the replay: https://t.co/Of9G2flEqD

[🤷 Nico Martin @nicodotdev]

Today marks one year at @huggingface 🤗

Transformers.js went from 150k to 3.7M weekly downloads, we shipped v4, Tokenizers.js, structured output and 207 WebGPU kernels.

I wrote down the whole story 👇
https://t.co/GCfj1nw8JW

[JSNation | The Main WebDev Conference @thejsnation]

🖐️ No clicks, no scripts — just an AI agent driving the browser itself. @nicodotdev, Machine Learning Engineer at @huggingface, built a multimodal AI agent that runs entirely client-side and holds context across conversations.

Catch the talk 👉 https://t.co/rpeAm4rV4u

[🤷 Nico Martin @nicodotdev]

I'll be back in SF for Web AI Summit!
A year after the v4 preview, Transformers.js went from 1.7M to 10M monthly downloads. Come see what shipped.
Also, @xenovacom will be talking about agents optimizing GPU kernels at web scale.

The whole WebAI world in one room: https://t.co/cQVT9mgofO

[Shreyas Karnik @KarnikShreyas]

GLiNER2.5-Decide now runs entirely in your browser 🎉

ONNX weights + a WebGPU demo in @nicodotdev's open-jev: 4 typed decisions in ~100 ms on a MacBook, and your text never leaves the tab.

Try it: https://t.co/52k03zT0fa

[🤷 Nico Martin @nicodotdev]

Try the demo yourself 👇
https://t.co/urCHFFByxY

[🤷 Nico Martin @nicodotdev]

Jev is the most exciting AI release in months. So I built the open, in-browser version 🎉

open-jev: System One style decisions running 100% on your device. Text in, typed answers out, one forward pass. Nothing is generated, so it can't hallucinate.

[🤷 Nico Martin @nicodotdev]

Under the hood: the KEV models by @jaredpalmer , open Apache-2.0 decision models with the same contract as @typesafeai 's Jev. Not the same model, but the same shape. I converted them to ONNX and they run on your GPU via WebGPU 🔥
https://t.co/eoYE0dmSB6