Meta Superintelligence Labs released Muse Glimmer on August 10, a 30-billion-parameter dense multimodal model distilled from the larger Muse Spark system and published under the permissive Apache 2.0 license on Hugging Face. The model pairs a 28-billion-parameter text decoder with a roughly 2-billion-parameter vision encoder, and 4-bit quantization compresses its memory footprint from about 55GB down to 18-20GB, letting the full stack, weights, KV cache, perception encoder, and speculative decoding drafter, fit inside a single 24GB or 32GB consumer GPU rather than a data-center card.
The model is purpose-built for always-on local agents rather than raw benchmark leadership: it supports a 131,000-plus token context window, more than 100 languages, persistent state across restarts, and failure recovery when a tool call breaks. On agentic benchmarks Meta reports it beating similarly sized open models Gemma4-31B and Qwen3.6-27B on MCP-Atlas, DeepSearch QA, and SWE-Bench Pro, though Qwen held the lead on computer-use and terminal tasks. Meta validated the release under its own Advanced AI Scaling Framework and states it does not meet the internal threshold for a frontier-risk model.
The release lands as Meta's fourth Muse-family model in five months, following Muse Spark in April, Muse Spark 1.1 in June, and the Muse Code coding agent earlier this month, and frames local, on-device agents as Meta's answer to closed, cloud-hosted rivals from OpenAI and Anthropic. If a 30B model can genuinely handle multi-step tool use and code editing on one card, it shifts part of the agent stack, and its per-token API costs, off cloud infrastructure and onto a user's own hardware.
