Muse Glimmer

Muse Glimmer is the model Meta Superintelligence Labs published on 10 August 2026, and its interest lies less in what it scores than in where it runs. It is a roughly 29.6-billion-parameter dense transformer distilled from the closed Muse Spark line — Meta says it was trained on the teacher's outputs by logit distillation on similar data — and released with full weights under a plain Apache 2.0 licence. That licence matters: Meta's earlier open releases carried a bespoke community licence with usage thresholds, and this one does not. The design target is stated plainly by the company: "always-on local agent workflows", meaning a model that sits on a laptop calling tools, writing and debugging code, reading screenshots and grinding through multi-step tasks without a round trip to a data centre. Vision comes from a frozen ViT-G/14 perception encoder of about 1.8 billion parameters bolted to the language model, and the context window is 131,072 tokens — a fraction of the million Muse Spark offers, which is the honest price of the size cut. Attention alternates three local layers, with a 2,048-token sliding window, to one global layer. The hardware story is the point. At BF16 the model wants 64 GB of memory, but Meta ships two 4-bit quantisations: one that fits 32 GB and costs 0.2% of quality, and one that fits 24 GB — a single RTX 5090 — for about 1%. Alongside the weights Meta released DFlash, a speculative-decoding drafter head, which it measures at 233 tokens per second on that card against 75 without it. Validation was done on RTX 5090 and MacBook M4-Max and M5-Max, and llama.cpp, Ollama, LM Studio and MLX integrations were announced with the model. Meta's published numbers put it at 76.0 on SWE-Bench Verified, 51.2 on SWE-Bench Pro, 65.9 on OSWorld-Verified, 94.7 on AIME 2026 and 83.5 on GPQA Diamond, with 75.4 on ScreenSpot Pro for screen work. All of these are the vendor's own figures from the model card; no independent index had scored the model at the time this profile was written. Meta states the release was assessed under its Advanced AI Scaling Framework, and publishes no external audit. Knowledge cutoff is 4 January 2026 and the model handles text and images in, text out, across more than a hundred languages.

▸News

Meta open-sources a distillation of its closed flagship: Muse Glimmer runs agents on one 24 GB card

8/11/2026 · Meta AI Research

▸Videos

No videos yet.

▸Reviews