Meta just shared the full recipe for a 30B agent model you can run at home 🧠

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while.

What the model is

Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards.

The parts I found most interesting

Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compute budget, the dense setups were 2 to 2.5 times more compute-efficient. Many local-AI conversations assume MoE always wins on small hardware, so it's good to see the trade-off measured.

Distillation all the way down. The base model learns from Muse Spark, Meta's much larger model, by matching its top 256 logits. For the first 95% of pre-training the loss is pure KL divergence. The last 5% switches to plain cross-entropy on the real tokens, to correct the bias you get from only seeing the teacher's top-k. After SFT, the student samples its own answers and the teacher scores them token by token (on-policy distillation), followed by RL. The authors say this beat an RL-only recipe at this size.

Synthetic "digital lives" for agent training. To teach tool use, they built simulated universes: a persona with 10 to 14 contacts, weeks of email, calendar events, documents and chat history, behind mock tools that return the same JSON as the real ones. The agent then gets realistic tasks with verifiable rubrics. It's a smart way to train on personal-assistant work without touching anyone's real data.

Quantization that barely costs anything. Their EvoPress mixed-precision profile shrinks the language-model weights to 19.65 GB and scores 75.6% on average over 15 tasks, against 75.7% for full BF16. A small DFlash drafter proposes blocks of 16 tokens for speculative decoding to speed up generation.

Keeping it honest

The benchmarks are strong but not a clean sweep. Against Gemma4-31B and Qwen3.6-27B, Glimmer leads clearly on MCPAtlas tool use (75.5 vs 54.2 and 62.5). Qwen3.6-27B is still ahead on SWE-Bench Verified (77.2 vs 76.0), OSWorld-Verified and TerminalBench 2.1. On a privacy test where lower is better, Glimmer (26.4) sits between Gemma (12.1) and Qwen (53.4). The report is open about this, which makes me trust the rest more.

What I like most is that the whole stack is public: full weights, 4-bit weights, a GGUF build, the drafter, and now the recipe. For homelab agents, that's a real gift. I'm curious to see how it holds up in real homelab setups. 🛠️

Sources

Reacties

Een reactie posten

Populaire posts van deze blog

SpaceX just asked to fly Starship hundreds of times a year 🚀

Last week the U.S. Federal Aviation Administration took a quiet but important step: it opened a formal review of SpaceX's request to launch Starship from Starbase in South Texas up to 456 times per year . Today the site is cleared for only 25 flights a year. That jump is the clearest signal yet that Starship is meant to become routine hardware, not a rare test rocket. What the FAA is reviewing On 14 October 2026 the FAA published a Notice of Intent to prepare a Programmatic Environmental Impact Statement (PEIS) for SpaceX's Starbase plans. A PEIS looks at a whole program of operations rather than a single launch, which is exactly what you need if you want a reusable rocket to fly the way airlines fly planes. SpaceX's proposal comes in two phases, according to the Federal Register notice: Phase 1: up to 100 launches and 200 landings per year (100 Ship, 100 Super Heavy), any day or time of week. Phase 2: up to 456 launches and 912 landings per year at Starbase. La...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

Google's Deep Review wants to be the second pair of eyes on every long document 🔍

AI has made writing fast. Checking has not kept up. That gap is exactly what Google Research is going after with Deep Review, which went public on 14 October. It's an agentic system that reads a long technical document and tries to find what's wrong with it: broken proofs, logic gaps, unsupported claims. For anyone working on legal and document AI, like me, this is a launch worth a close look. From conference pilot to public tool Deep Review is the grown-up version of Google's Paper Assistant Tool (PAT). Over the past year, authors of more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027 opted in for private feedback before the submission deadline, fully separate from official peer review. The need is real: Google counts over 74,000 submissions to ICLR, NeurIPS and ICML this year, up from about 23,000 in 2023. Now anyone can use it. Every user gets two free reviews per week, and you can connect a Gemini API key for higher limits and programmati...
Meer lezen