Meta ships Muse Glimmer, a 30B open weight model built to run agent loops on a single GPU

8:00 AM PT · August 11, 2026

Meta’s Muse Glimmer is a dense, 30 billion parameter model built around the loop an autonomous agent actually runs: plan, call a tool, read the result, keep going, and recover cleanly when something breaks. Released under a permissive Apache 2.0 license, it was compressed to roughly 4 bit precision and paired with block level speculative decoding so it stays fast enough to sit inside a live agent session, fitting within a 20 to 32 gigabyte memory envelope on hardware like an RTX 5090 or a MacBook M4 or M5 Max. Meta says Glimmer outperforms similarly sized open models, including Gemma4-31B and Qwen3.6-27B, on agentic, coding, multimodal, safety, and reasoning benchmarks such as DeepSearch QA, MCP-Atlas, and SWE-Bench, and it supports more than 100 languages with a context window above 131,000 tokens. The release lands alongside CEO Mark Zuckerberg’s essay making the case for distributing capable AI broadly rather than concentrating it in a handful of labs, and Meta says open weights for its larger Muse Spark 1.2 model are coming next. Running fully on device, without a network call, is the core pitch: an agent that manages a calendar, drafts messages, or organizes files without routing anything through the cloud.