Meta released Muse Glimmer on August 10, 2026, a 30-billion-parameter open-weight model designed to run entirely on consumer hardware including Macs, PCs, and single GPUs. The model is optimized for local, always-on agent workflows—performing multi-step agentic tasks, tool use, scheduling, file organization, and code generation with latency low enough to maintain workflow continuity. Meta published the weights under an Apache 2.0 permissive license on Hugging Face, enabling immediate deployment and customization.
Muse Glimmer distills capabilities from Meta's larger Muse Spark 1.2 teacher model through a novel multi-stage training recipe (pre-training logit distillation, mid-training longer-context data, post-training supervised fine-tuning plus on-policy distillation and reinforcement learning). The model compresses to under 20GB via quantization and includes a lightweight DFlash speculative decoding drafter to reduce generation latency. Benchmarks show it delivers competitive agentic performance: top-5 results on DeepSearch QA (74.6), SWE-Bench Verified (76.0), and GPQA Diamond (83.5), beating Gemma4-31B and Qwen3.6-27B on several key agentic tasks.
Meta also announced it will release open-weight versions of Muse Spark 1.2, its most capable model, in the coming weeks. Day-0 integrations landed in transformers, llama.cpp, vLLM, MLX, and ExecuTorch. This release marks Meta's pivot toward supporting local, privacy-preserving deployments as enterprises grow wary of ballooning cloud AI costs and centralized dependencies, while Zuckerberg has positioned open-weight models as critical to competition with closed vendors and international rivals.