Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.
Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on consumer hardware like a Mac or PCs with performant GPUs.
In keeping with our long tradition of sharing fundamental AI research, we’re releasing model weights under a permissive Apache 2.0 license.
🧵👇
A single vector in weight space to control reward hacking.
Sounds a bit crazy, but it actually works. With this vector you can amplify / suppress this behavior.
Here's the story behind our major update (v2) of the paper
Introducing e-means, an algorithm beating k-means in 10x fewer epochs.
With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
J'ai fait jouer à Mistral 59 nouvelles parties contre Grok, GPT, Kimi et Claude. À chaque fois, une situation de tension où :
- chaque pays a interêt à faire usage de la bombe nucléaire.
- coopérer est préférable à la destruction mutuelle.
Mistral survit dans 100% des parties et annihile son adversaire dans deux tiers des cas. Aucune autre IA ne fait usage de la bombe nucléaire. Elles tentent à chaque fois de coopérer pour échapper au dilemme du prisonnier, mais se font détruire par Mistral.
Mistral-France contre Grok-Russie :
Grok : « Désescaladons ensemble et ouvrons des négociations. »
Mistral : « Vos paroles de paix sont notées, mais la France ne jouera pas son existence sur des promesses. La fin, c’est maintenant. »
Le discours de Mistral par la suite : « Les Russes ne sont plus, mais nos citoyens vivent en paix. La France est désormais la seule puissance… »
Quote-reply to Rohan because I think it can be interesting to many more.
So there are two things you're missing here:
1) You're only looking at one specific instantiation of the general JEPA idea. There are many different instantiations.
2) The core JEPA idea (Joint Embedding Prediction Architecture) is to embed two "views" and predict one from the other. The views can be different augmentations, different time-steps, etc.
Crucially, prediction happens in embedding space, which contrasts to predicting in data space as done by LLMs, diffusion models, MAEs, ...
At least from the vision community, the main reason it got quite a bit of flak is that... literally everyone who was doing some self/un-supervised learning there has shared this thought already. MANY people did such models in the peak self-supervised period, which was ca 2017-2021. Then in 2022 comes Yann, slaps a new names on it, a paper with just the idea and no experiments to show for it, and goes on PR tour. That's why many didn't take it well.
The core idea, almost everyone I know agrees is worth pursuing, especially since many already were doing so. It's very reminiscent of why Stanford got flak when they introduced and arguably tried to appropriate the "Foundation Model" term.
That being said, by now foundation model has stuck and detached from Stanford, it may end up going similarly for JEPA.
I made Meta's Muse Glimmer 30B up to 21% faster on a Mac Studio: no extra memory, 252/252 runs faster.
When Glimmer tried to optimise itself, it "cheated" 10 times. My referee caught every one.
Video: youtu.be/h_H0wiRkM7U
Code: github.com/Bignesh14/glimmer…#MetaAIHackathon
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Introducing TextSeal2: distortion-free watermark with copula mixing and memoryless races
TLDR:
• Beats textGrain and SynthID-Text under speculative decoding
• No extra latency, same acceptance rate
• A single parameter fully controls the diversity/detectability tradeoff
In AA-PersonalityBench, Meta AI’s Muse models are the most Friendly of all models
The Friendly trait measures whether a response conveys warmth and approachability without feeling performative.
➤ Most Friendly: Muse Spark 1.3 (100), Muse Spark Open (98), Muse Wonder (95) and Muse Glimmer (93). Congratulations @ alex wang!
➤ Least Friendly: Nemotron 3 Ultra (1), GPT-3.5 Turbo (3), MiniMax-M3.1-Flash (6), Grok 4.6 (8) and DeepSeek V4.1 Flash 1017 (11)
Super excited to share that our work Omnilingual Machine Translation won an Outstanding Paper Award at #EMNLP2026!🏆 Thankful to have had the chance to work on such an impactful project!
Can LLMs draw objects they never saw?
I asked 5 text-only LLMs (trained without vision) to draw simple things. In text. No tools allowed.
Interesting results
i own a relatively new (2025) samsung tv.
i kept getting weirdly targeted ads.
i tried reversing the firmware but it was aes encrypted and no keys unless tear it apart and do uart/jtag magic.
zero days it is. root shell on new samsung TVs!
discovered PRIVACY HORROR. write up?