Meta took the wraps off Muse Glimmer on Monday, a 30B multimodal model that runs agentic workloads on a single consumer GPU. The open-weight release, distilled from the company’s larger Muse Spark system, ships under an Apache 2.0 license and comes pre-quantized with a multi-token-prediction drafter so it responds quickly enough for interactive use on a Mac or PC.
Benchmarks released by Meta put the model ahead of comparable open rivals in most agent categories. On SWE-Bench Pro, Glimmer reached 51.2, beating Google’s Gemma4-31B at 36.9 and edging Alibaba’s Qwen3.6-27B at 50.2. Its 75.5 on the MCP Atlas test for general agentic skill clears both competitors by a wide margin. The model also works with OpenClaw and other open agent frameworks without extra setup.
CEO Mark Zuckerberg framed the release as part of a broader open-source campaign, warning that advanced AI should not end up controlled by a handful of companies and asking US officials to lower barriers for open-weight development. Meta has leaned on open releases to counter the closed frontier systems sold by OpenAI and Anthropic.
Glimmer lands as the open-weight segment fills up fast. Tencent shipped its 295B Hy3 under Apache 2.0 in recent weeks, and smaller agent-ready models from Liquid AI and Mistral target the same developers. Meta’s wager is that a capable model small enough for one GPU becomes the default choice for local agent work.
