Searching for something new to read? All ICML 2026 papers are now public: openreview.net/group?id=ICM... (including all 6551 accepted papers and also rejects which opt-in)

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.

Three new model weights: 0.5B, 1.5B, 4B are available using new quantization and distillation techniques. Download #Apertus 1.1, read the ICML workshop report, try a new demo on @hf.co - all just a tap away in our latest blog post: apertus-ai.org/articles/202...

Announcing the #ICML2026 invited speakers! Pascale Fung Susan Athey (@susanathey.bsky.social) Sham Kakade (@shamkakade.bsky.social ) Aviv Regev Verena Rieser (@verenarieser.bsky.social) Arvind Narayanan (@randomwalker.bsky.social) Check out the blog post for more info! blog.icml.cc/2026/05/18/a...

Ahem, back to business... Decision notifications are being released on OpenReview. There were 23,918 submissions that entered review, roughly double last year. 6,352 papers were accepted, for an acceptance rate of 26.6%. 536 papers (2.2% of submissions) are "spotlights." 1/3

Muon: I made a new 3-slides explanation of this amazing optimizer for today's lecture. Let me know what you think

Every system that was regulated, either explicitly or implicitly, by the fact that they were effortful for humans (letters of recommendation, government filings, essays, or, as this paper finds, lawsuits) will break under a wave of AI.

To ensure compliance w peer-review policies, ICML has removed 795 reviews (1% of total) by reviewers who used LLMs when they explicitly agreed to not. Consequently, 497 papers (2% of all submissions) of these (reciprocal) reviewers have been desk rejected Details in blog post 👇

There has been some online discussion of prompt watermarks in ICML submissions. tl;dr: - Yes, this is one of the *conference*'s (several) scientific integrity measures - Yes, it's not infalliable (but it still helps) - No, your paper won't be desk rejected as a result 1/4

More like every week.

Open models continue to pace closed models on a 9 month lag.

A factor of 10 billion since 2010 😮 A couple of eye-opening slides form @sloeschcke.bsky.social's presentation at today’s @belongielab.org meeting (1/2)

The #ICML2026 abstract deadline has passed! We're at 33540 active abstracts (and dropping). How many will make it over the finish line? 🏁

New blog post (on a shiny new ICML blog!): What's New in #ICML2026 Peer Review Some highlights: - Policies to combat thinly sliced contributions - Cascading desk rejections for peer-review abuse - Reviewer reciprocity - New ways to support authors and reviewers Post: blog.icml.cc/2026/01/08/w...

A multidisciplinary team of ETH Zurich researchers developed a method of using an autonomous excavator to construct a dry-​stone wall that is six metres high and sixty-​five metres long.

We updated the plots we use to measure the open model ecosystem at interconnects, to guide The ATOM Project, and to understand what's happening. We have ~8 plots to summarize what's happening. First, the high level picture showing China's growing adoption lead.

Announcing the ICML 2026 policy for LLMs in reviewing! Reviewers and authors both pick either conservative or permissive LLM use, and will be matched accordingly. Importantly: authors on papers who choose conservative must obey the conservative policy as reviewers.

👀 I am working on something pretty cool.. Hopefully, it will soon be possible to try #Apertus 🇨🇭 directly in your browser, powered by Transformers.js 🎉

The threshold for consistent English/query understanding is now 3M parameters.

Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...

🎉 ICML 2026 Call for Papers (& Position Papers) is here! 🎉 📅 Key Dates Abstract deadline: Jan 23, 2026 AOE Paper deadline: Jan 28, 2026 AOE A few key changes this year: - Attendance for authors of accepted papers is optional - Originally submitted version of accepted papers will be made public ...

so open-weights models are much happier than closed ones i guess, cause they live on in the long run, did i get that right?

91% of reasoning does not need RL 🤯 arxiv.org/abs/2510.07364

I just tried the official demo for the new Gemini 2.5 Computer Use model and it started by navigating to Google, solving Google's own CAPTCHA and then running a search! https://simonwillison.net/2025/Oct/7/gemini-25-computer-use-captchas/

We're hiring again for AI research engineering roles: Join the team behind the Apertus LLM, if you share our passion to work on impactful AI that's truly open. careers.epfl.ch/job/Lausanne...

1/🚨 New preprint How do #LLMs’ inner features change as they train? Using #crosscoders + a new causal metric, we map when features appear, strengthen, or fade across checkpoints—opening a new lens on training dynamics beyond loss curves & benchmarks. #interpretability

Schweizer Sprachmodell Apertus: So sieht EU-konforme, transparente KI aus https://www.heise.de/hintergrund/Schweizer-Sprachmodell-Apertus-So-sieht-EU-konforme-transparente-KI-aus-10638501.html?utm_source=flipboard&utm_medium=activitypub Gepostet in Nachrichten @nachrichten-heiseonline

Hugging Face's FinePDFs The largest publicly available corpus sourced exclusively from PDFs, containing about 3 trillion tokens across 475 million documents in 1733 languages. - Long context - 3T tokens from high-demand domains like legal and science. - Heavily improves over SoTA

you can run the new apertus LLMs fully locally on your (mac) laptop with just 2 lines of code: pip install mlx-lm mlx_lm.generate --model swiss-ai/Apertus-8B-Instruct-2509 --prompt "wer bisch du?" (make sure you have done huggingface-cli login before)

Am Schluss müssen sich die Medienverlage eine gesonderte Lösung überlegen, da sie kaum für alle Schweizer Blogger, Firmenwebsites, Künstler:innen,Gesundheitsportalen, eCommerce-Plattformen sprechen können. WBK N will weder Opt Out noch Opt In festschreiben.

The next generation of open LLMs should be inclusive, compliant, and multilingual by design. That’s why we @icepfl.bsky.social @ethz.ch @cscsch.bsky.social ) built Apertus.

new extensive evaluation of different optimizers for LLM training arxiv.org/abs/2509.01440

Die Schweiz steigt ins Rennen der grossen Sprachmodelle ein. Unter dem Namen #Apertus veröffentlichen @ethz.ch, @icepfl.bsky.social und das @cscsch.bsky.social das erste vollständig offene, mehrsprachige #LLM des Landes. Fürs MAZ habe ich Apertus kurz analysiert: www.maz.ch/news/apertus...

EPFL, ETH Zurich & CSCS just released Apertus, Switzerland’s first fully open-source large language model. Trained on 15T tokens in 1,000+ languages, it’s built for transparency, responsibility & the public good. Read more: actu.epfl.ch/news/apertus...

EPFL and ETH Zürich are building together a Swiss made LLM from scratch. Fully open and multilingual, the model is trained on CSCS's supercomputer "Alps" and supports sovereign, transparent, and responsible AI in Switzerland and beyond. Read more here: ai.epfl.ch/a-language-m... #ResponsibleAI

Why did Grok suddenly start talking about “white genocide in South Africa” even if asked about baseball or cute dogs? Because someone at Musk’s xAi deliberately did this, and we only found out because they were clumsy. My piece on the real dangers of AI. Gift link: www.nytimes.com/2025/05/17/o...

If you’re at @iclr-conf.bsky.social this week, come check out our spotlight poster INCLUDE during the Thursday 3:00–5:30pm session! I will be there to chat about all things multilingual & multicultural evaluation. Feel free to reach out anytime during the conference. I’d love to connect!

Using the 'right' data can hugely speed up LLM training, but how to find the best training data in the vast sea of a whole web crawl? We propose a simple classifier-based selection, enabling multilingual LLMs 🧵

Dion: A Communication-Efficient Optimizer for Large Models (inspired by PowerSGD) arxiv.org/abs/2504.05295

Prime Intellect's INTELLECT-2 The first decentralized 32B-parameter RL training run open to join for anyone with compute — fully permissionless. www.primeintellect.ai/blog/intelle...

Anastasia @koloskova.bsky.social recently won the European @ellis.eu PhD award, for her amazing work on AI and optimization. She will be joining University of Zurich as a professor this summer, and hiring PhD students and postdocs. You should apply to her group! Her website: koloskova.github.io

The Swiss AI Initiative has launched open calls for disruptive ideas - Democratizing large-scale AI for the benefit of society. Send your idea by end of March 🏃‍♂️‍➡️ , and run on one of the largest public AI clusters globally. Everyone is eligible to apply! swiss-ai.org

excellent starting point for pretraining recipes: arxiv.org/abs/2502.02737

new open weights, 24B model, with comparable performance to Llama 3.3 70B 😮. congrats mistral team! mistral.ai/news/mistral...