Meta's Muse Glimmer 30B Runs AI Agents on a Single Consumer GPU

Meta released Muse Glimmer, a 30B open-weight model built for agentic tasks that runs at practical speeds on a single consumer GPU at 4-bit precision, on a Mac or PC. The model is distilled from Meta's larger Muse Spark model and trained on synthetic agent runs in mock applications to improve performance on long tool-using tasks. Its 4-bit quantized weights fit under 20 GB, making it runnable on a single RTX 5090. It handles time-based and environment-driven events mid-episode, including incoming messages, state changes, and scheduled triggers. Meta published a technical report covering model architecture, distillation recipe, and quantization.

Across benchmarks, performance is competitive but uneven. On MCPAtlas, a multi-tool benchmark, Muse Glimmer scored 75.5, compared to 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. On SWE-Bench Verified at the high compute tier, it scored 76.0, ahead of Gemma4-31B (66.6) and roughly matching Qwen3.6-27B (77.2). On WildClawBench it scored 47.6, ahead of Gemma4-31B (37.6) and Qwen3.6-27B (43.2). On OSWorld-Verified, Muse Glimmer trails: it scores between 58.7 and 69.7 depending on compute tier, versus Qwen3.6-27B's 75.6. The evaluation uses 800 tasks across five capability splits covering execution, search, ambiguity, adaptability, and time, using the OpenClaw harness.