Meta's Muse Glimmer brings local AI agents to consumer GPUs

Meta's Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter open model built specifically to run AI agents on a user's own machine rather than in the cloud. The model is licensed under Apache 2.0, and its weights are already available on Hugging Face.

The pitch is privacy paired with capability. Muse Glimmer is designed to handle everyday personal tasks such as managing calendars, sorting files, and reading messages entirely on-device, so sensitive data never has to leave the user's hardware.

That local-first ambition depends on the model actually fitting on consumer machines. At full precision, Muse Glimmer needs more than 55 GB of memory, but 4-bit quantization shrinks that footprint to under 20 GB, putting it within reach of high-end gaming GPUs and MacBooks. Meta also pairs it with a small helper model that it says speeds up text generation by as much as 3.1x.

Muse Glimmer was created by distilling Meta's larger Muse Spark model, and it supports agent-orchestration frameworks including OpenClaw alongside custom scaffolds, positioning it for developers building autonomous task-execution pipelines rather than chat interfaces.

How it stacks up against rivals

Meta benchmarked Muse Glimmer against two other open models in its weight class, Gemma4-31B and Qwen3.6-27B. The results are mixed rather than a clean sweep, which is itself notable for a model this size.

Muse Glimmer leads on several agent-centric tests: 23.5 on τ2-Banking versus 15.1 and 16.7 for its rivals, 75.5 on MCP Atlas versus 54.2 and 62.5, and 47.6 on WildClawBench versus 37.6 and 43.2. It also edges ahead on GAIA2 (43.3), SciCode (43.6), and DeepSearch QA (74.6).

But Qwen3.6-27B still wins on several coding and execution benchmarks, including SWE-Bench Verified (77.2 vs 76.0), SWE-Bench Pro (50.2 vs 51.2, where Glimmer actually edges ahead), OSWorld-Verified (75.6 vs 65.9), TerminalBench 2.1 (60.7 vs 51.7), and SkillsBench (46.6 vs 44.3). The picture is a competitive open-weight field rather than a decisive win for any single model.

Context: a strategy reset at Meta

The release lands alongside a broader repositioning of Meta's AI research. Superintelligence Labs is the unit that emerged from a reorganization of Meta's AI group following Chief Scientist Yann LeCun's departure, and it is now the home for Meta's open-model efforts.

Mark Zuckerberg used a companion essay, titled "The Future is for Everyone," to defend that open strategy, including Meta's practice of distilling techniques from other labs' models into its own releases. The framing suggests Meta sees openly licensed, locally deployable models as a competitive answer to closed rivals and to fast-moving open models coming out of China.

For developers, the practical takeaway is a genuinely deployable agent model: small enough to quantize onto a single consumer GPU, permissively licensed, and tuned for the kinds of orchestration and tool-use benchmarks that matter for autonomous agents rather than pure chat quality.