Meta released Muse Glimmer (also called Muse Glimmer-30B), an open-weight ~30-billion-parameter dense multimodal model optimized for local/on-device agentic workflows, on August 10, 2026. It comes from Meta Superintelligence Labs (MSL) and is distilled from their larger Muse Spark foundation model series.
Mark Zuckerberg announced the release on X, stating that Meta is opening the weights for Muse Glimmer (a 30B dense model that can run locally) and will soon release weights for Muse Spark 1.2 (their latest foundation model). He emphasized Meta’s support for open source. AI at Meta and other accounts posted additional details and demos the same day.
It is purpose-built for autonomous/local agent tasks rather than pure chat: multi-step reasoning, reliable tool/function calling with schema adherence, failure recovery (diagnose and retry), long-horizon planning, and end-to-end task completion.
Full-precision would require 55+ GB. Meta quantizes weights to ~4-bit (e.g., K-Quant variants), bringing the language model under 20 GB. This leaves headroom in a 24 GB or 32 GB VRAM envelope for the KV cache, perception encoder, and speculative decoding drafter. Degradation on agentic tasks is described as minimal to none.
It ships with a lightweight DFlash-based speculative decoding drafter (block proposals verified in parallel by the main model) for substantial speedups while preserving output quality:
It is designed to run fluidly for conversation and real-time agent interaction entirely offline on a Mac or PC with a single performant consumer GPU.
Muse Glimmer shows strong results versus similarly sized open models (e.g., Gemma 4 31B Thinking and Qwen3.6 27B Thinking modes) on agentic, coding, multimodal, and reasoning benchmarks. Notable wins or competitive scores appear on MCP-Atlas, DeepSearch QA, τ3-Bench variants, WildClawBench, SWE-Bench Pro, certain multimodal and high-reasoning suites (AIME 2026, etc.). Full tables and methodology are in the model card and related reports.
Controllable reasoning effort (via system prompt settings such as high/xhigh) allows quality/speed trade-offs.
meta-models/Muse-Glimmer-30B).Demos highlight end-to-end agent behavior (e.g., discovering a local Home Assistant instance, querying APIs, generating and deploying a dashboard from a natural-language prompt).
This marks a notable return to strong open-weight releases after earlier Muse Spark models (which started closed). It positions Meta in the local/edge agent race against models like Gemma variants and efficient Chinese open models, while the forthcoming Muse Spark 1.2 open weights aim higher on the capability ladder. Zuckerberg has also publicly advocated for reduced U.S. policy friction on open-source AI training data to better compete internationally.
Because the release is brand-new (August 10, 2026), community fine-tunes, additional quantizations, detailed independent benchmarks, and broader tool integrations are still rolling out. Check the Hugging Face page and Meta’s developer resources for the absolute latest files, quant variants, and docs.
