Muse Glimmer is a 30-billion-parameter (29.6B including vision encoder) dense causal language model with a dedicated ~1.8B-parameter ViT-G/14 perception encoder, released by Meta Superintelligence Lab as a distillation of Muse Spark purpose-built for autonomous agentic tasks running entirely on consumer hardware without cloud infrastructure. It integrates multi-step reasoning, reliable schema-based tool use, multimodal understanding (interleaved text and images), and failure recovery into a single locally-runnable model, with a 131K+ token context window, controllable reasoning strength (low/medium/high/xhigh), and compatibility with agentic scaffolds like OpenClaw and Hermes Agent. Optimized for on-device deployment, 4-bit quantization shrinks the model under 20GB with minimal degradation (0.2–1.0% across 15 benchmarks), and it ships with a DFlash speculative-decoding drafter that predicts 16-token blocks for up to 3.1x generation speedup on an RTX 5090 (233 tok/s) and 1.8x on an Apple M5 Max (50 tok/s). On agentic and reasoning benchmarks, Muse Glimmer generally outperforms similarly-sized Gemma4-31B and often trades blows with or edges out Qwen3.6-27B — leading on MCP Atlas (75.5), DeepSearch QA (74.6), Gaia2 (43.3), and AIME 2026 (94.7) — while trailing on tasks like OSWorld-Verified and TerminalBench 2.1; it was assessed by Meta's Preparedness Team at "moderate or lower" risk across chem/bio, cyber, and loss-of-control axes, trained with safety SFT and RL for tool-use boundaries and prompt-injection resistance, and is released under Apache 2.0 with full-precision, two 4-bit quantized variants, the DFlash drafter, and the perception encoder all publicly available; it is not intended for use by individuals under 18.
Limitations from Meta
The model may produce inaccurate, biased, or objectionable responses to user prompts.
While optimized for agentic tasks, the model may still make errors in multi-step reasoning, particularly in novel scenarios not well represented in training data.
The model is not explicitly optimized for video; video input is processed as individual frames.
The model has not been evaluated on all languages contained in the pre-training data. Performance may degrade on languages outside the strongly supported set.
Quantized inference may show minor quality differences in edge cases compared to full-precision.
The model is not intended to be downloaded by or used by individuals under the age of 18. Where deployed within systems that may be used by individuals under the age of 18, deployers are responsible for ensuring that any risks associated with such use by individuals under the age of 18 has been fully assessed and appropriately mitigated, and complies with all applicable laws.
Responsible Use: Developers should perform their own safety testing and tuning tailored to their specific applications and proposed languages. Our Usage Policy can be found here [link]. We recommend implementing additional guardrails (such as human-in-the-loop confirmation for irreversible actions) when deploying the model in agentic contexts where it can take real-world actions.
