🤷 Nico Martin

Apertus Mini is now running entirely in your browser 🇨🇭 80+ tps for the 1.5B, 60+ tps for the 4B (on my M3). Fully client-side via Transformers.js + ONNX + WebGPU.

Apertus isn't just an LLM. it's the Swiss AI Initiative: @epfl-ai-center.bsky.social , @ethz.ch and @cscsch.bsky.social building open weights, open data, open science. And truly open should run anywhere. Including on your device. So as a Swiss dev, I had to bring it to the browser.

Apertus Mini is now running entirely in your browser 🇨🇭 80+ tps for the 1.5B, 60+ tps for the 4B (on my M3). Fully client-side via Transformers.js + ONNX + WebGPU.

On my way to San Francisco🤩. First time in California. First time in Silicon Valley. And I'm going for Google I/O. 🤯 Also I recently started a new instagram channel. So if you want the unfiltered version, follow along👇😊

My learning from Uphill Conf '26: being a speaker is so much easier than being an organizer 😅 After the last two years on stage, I switched sides. Nine months of prep, a lot of group chats, and one very humbling ropes course at the speaker event later, we pulled it off!

And my on-stage-contribution this year? I got to moderate the open source panel with @benburtenshaw.bsky.social, @jonny.io, Panos Vagenas and @rjourdan.com.

My learning from Uphill Conf '26: being a speaker is so much easier than being an organizer 😅 After the last two years on stage, I switched sides. Nine months of prep, a lot of group chats, and one very humbling ropes course at the speaker event later, we pulled it off!

@osanseviero.bsky.social walked through what Gemma4 means for on-device AI. New benchmarks, solid technical depth.

My learning from Uphill Conf '26: being a speaker is so much easier than being an organizer 😅 After the last two years on stage, I switched sides. Nine months of prep, a lot of group chats, and one very humbling ropes course at the speaker event later, we pulled it off!

over 300 attendees enjoyed a sunny and inspiring day on the Gurten. This year with a second stage and a focus on valuable insights, optimized solutions, and lessons learned from running AI apps in production to share with the developer community. A few highlights from the talks:

Getting started with AI in the browser has never been easier 🤗 Today we're releasing the Transformers.js #AgentSkill, your shortcut to running ML models in the browser. It provides all the information your coding agent needs to add on-device AI to your web app.

TranslateGemma 4B by Google now runs 100% in your browser on WebGPU with Transformers.js v4. 55 languages. No server. No data leaks. Works offline. A 4B parameter translation powerhouse, right in your browser. Try the demo 👇

Exciting news: this year I'm joining the organizing team of @uphillconf for the first time! 🎉 For the past two years, I've had the opportunity to give a talk, and each time it was a fantastic experience. That's why I decided to actively participate in the planning this year😊

I've seen several posts lately that share some edge cases in AI conversations where the AI makes mistakes. Of course, this is to show how bad AI still is. AI is a tool. If it produces incorrect or unusable results, it is not the tool that is at blame, but the person who is using it incorrectly..

🤗 Transformers.js v4 preview is now available! npm i @huggingface/transformers@next And its huuuuge! I don't think a single line of code remained untouched😄. New build system, monorepo structure, and tons of new features!

Over the Christmas holidays, I built three apps to replace my Google Sheets workflows: - 𝗙𝗶𝗻𝘁𝗿𝗮𝗰𝗸: A budget tracker that imports bank transactions and allows us to categorize them - 𝗣𝘂𝗺𝗽𝗲𝗹: A gym tracker - 𝗛𝗮𝗯𝗶𝘁𝘀: A daily habit tracker

The lesson: AI coding tools work best when you establish clear structure upfront. The time you invest in architecture pays dividends in maintainability and makes AI assistance more effective, not less.

Over the Christmas holidays, I built three apps to replace my Google Sheets workflows: - 𝗙𝗶𝗻𝘁𝗿𝗮𝗰𝗸: A budget tracker that imports bank transactions and allows us to categorize them - 𝗣𝘂𝗺𝗽𝗲𝗹: A gym tracker - 𝗛𝗮𝗯𝗶𝘁𝘀: A daily habit tracker

Components bloat, there's no clear structure for sharing logic or components, and the codebase becomes increasingly fragile.

Over the Christmas holidays, I built three apps to replace my Google Sheets workflows: - 𝗙𝗶𝗻𝘁𝗿𝗮𝗰𝗸: A budget tracker that imports bank transactions and allows us to categorize them - 𝗣𝘂𝗺𝗽𝗲𝗹: A gym tracker - 𝗛𝗮𝗯𝗶𝘁𝘀: A daily habit tracker

I approached each project with different levels of "vibe coding." For Fintrack, I wrote most code myself or carefully reviewed every line. I wanted to understand exactly where data is stored and how it flows through the system (for obvious reasons 😅).

Over the Christmas holidays, I built three apps to replace my Google Sheets workflows: - 𝗙𝗶𝗻𝘁𝗿𝗮𝗰𝗸: A budget tracker that imports bank transactions and allows us to categorize them - 𝗣𝘂𝗺𝗽𝗲𝗹: A gym tracker - 𝗛𝗮𝗯𝗶𝘁𝘀: A daily habit tracker

Definitely not an easy post to write. A friend of mine died in an avalanche a few days ago. He was a professional snowboarder who knew the mountains intimately.

When I was younger, I often went freeriding. Skiing in deep powder is one of the best feelings there is. But when our first child was born, I gave away my powder skis and never went in deep powder again.

I made myself a little Christmas present: a workout tracker that runs 100% offline in your browser. I's built entirely with Claude Code, zero manual coding, just conversations. Every component, every database query, even the deployment config. pumpel.nico.dev