Happy to release ✨ Muse Glimmer ✨ - level ~= Qwen 3.6-27B - Apache 2 - 30B dense - quantized to run in 17GB - quant + spec dec => 50 tok/s on macbook m5 max, interactive, smooth Enjoy!

"the not laughing cow" smh what a pale copy of the original

this is actually just straight up good shit. genuine art. what is happening

2000-2010: cherished memories 2010-2015: The Minecraft Black Hole 2015-2020: not gaming (???) 2020-now: beautiful indies (top-left is melee)

Hot damn! Congrats on the AGI, @OpenAI! 😉 kenforthewin.github.io/blog/…

Achh Lucas beats me but also I physically cannot have a lower slop index with this number of papers so I'm happy lol

🔥 PhD in the DINO team 🔥 We have a PhD program opening in our team. Don't hesitate to reach out to me or Max if interested!

The DINO team at FAIR Paris is hiring a PhD student to push the frontier of visual representation learning. We offer hard research topics, a research and engineering team supporting you, and a stimulating PhD community. Help us shape the next generation of DINO models! Details👇

> "Geoff, get a life" Fucking ouch

I never understood the whole encoder/decoder naming in transformers. Everything is just self attention or cross attention with different attention masks with no typical volume change. The enc / dec naming made sense in VAEs where you bottleneck and learn appropriate posteriors.

there was a time when you made your last matplotlib figure and you didn't know it was the last one

Want to bring this pet peeve to light: "decoder-only" transformer has never been a decoder, and never will be. It goes from data space to data space. If you want to say it's causal, just say it's causal

Chinchilla writes loss as a sum of two independent terms: model size and data never interact. Kaplan's original law coupled them. But interaction has an impact, dropping it bends predictions exactly where you extrapolate. Skaling, our new law, puts it back with one exponent. 1/8

Meta AI researcher who created DINOv2 explains modern self-supervised learning! Also introduces a new approach called CAPI @TimDarcet gave a brilliant talk about his research at the @MedARC_AI journal club last month. Very clear and information-dense, I learned a lot from his talk. Give it a watch!

✨Excited to share the release of Muse Glimmer, a 30B dense open-weight model optimised for local agentic workflows (🧵) Muse Glimmer: - is built for end-to-end agentic task completion, combining reliable tool use, multi-step reasoning, failure recovery, and compatibility with agentic scaffolds to solve complex tasks from start to finish - offers multimodal reasoning and controllable effort, enabling it to interpret text, images, charts, and documents while balancing reasoning depth with speed - is tailored for local deployment to empower more and more people to realise their goals and aspirations 📊Blog post research.meta.ai/blog/introd… 🤗Huggingface huggingface.co/meta-models/M… 📚Developer Documentation dev.meta.ai/docs/muse-glimme… 🌱 Mark’s essay on AI meta.com/thefutureisforevery… Truly grateful to my dear colleagues for the rollercoaster ride Antoine Grosnit, Alexis Audran-Reiss, Arash Abtahizadeh, Adil Ahmed, @AsafAvrahamy, Svetlana Kurtsman, @LuciaCKun, Muna Aghamelu, Noam Levi, He Ye, Joy Chen, Gal Cohen, Onur Çelebi, @roccajo, @HugoTouvron, @CheeHauLeow, Igor Tufanov, Eduardo Sánchez, @PereLluisHC , Hela Momand, @BhavulGauri, Parth Pathak, Kelvin Niu, @kumarsujan, Anton Protopopov, Gaurav Chaurasia, @albertomariape , Jade Yu, @GagnonAudet, @_masoudjalili, @arreqe_ai, @ishitamed, Mattia Opper, Paris Giampouras, @rrmaura , @TimDarcet , Saba Nazir, Sandra Lefdal, @alex_h_miller , @rybolos , Luca Sbordone, Grant Gardner, @maryamfazel and, of course, the one and only Yoram Bacrach! Big thanks to the Meta AI leadership @NailaMurray , @rob_fergus , @LukeZettlemoyer , @skasriel @alexandr_wang , @finkd for making this happen 🙏

Muse Glimmer, an open-weight 30B agentic model that runs locally. It’s been incredibly fun to be a part of this journey and work alongside the amazing MSL team! ✨ Excited to see what the community builds with it! 👇 🤗: huggingface.co/meta-models/M… | 📝: research.meta.ai/blog/introd…

@AIatMeta releases new open source LLMs! Exciting stuff!

A year ago I was hoping for this !! Today it finally happened It may not be called Llama anymore, but seeing Meta release a frontier open weight model under an Apache license feels like history repeating itself. Huge win for the open ecosystem.

Llama 5 basically. Super excited

Apache 2.0 🥹 first Gemma, now Meta we are so back

Meet Muse Glimmer: an open-weight model built for always-on local agents. 30B parameters, Apache 2.0 and tuned for complex multi-step work so it plans, calls tools, hits errors, retries, and sees the task through long-horizon loops. Muse Glimmer is designed to balance capability against the memory and compute constraints of local hardware. We couldn’t be more excited to get this in the hands of developers. Download and start building now: bit.ly/4yYUKkS