The announcement
On August 10, 2026, Meta Superintelligence Labs introduced Muse Glimmer, a roughly 30-billion-parameter model with weights under Apache 2.0 on Hugging Face. The focus is always-on local agents: function calling, local coding, multimodal (text and image) and LLM-as-a-judge.
How it fits consumer hardware
At full precision, 30B would need over 55 GB. Meta uses ~4-bit quantization to fit under 20 GB, with room for KV cache, a perception encoder and a speculative decoding drafter (DFlash). Speed demos covered MacBook M4/M5 Max and RTX 5090.
Training and availability
The model was distilled from Muse Spark outputs, then mid-trained and post-trained (SFT + RL) on agentic tasks. Integrations with llama.cpp, MLX, ExecuTorch, Ollama and LM Studio were expected in the following days. For SMEs, the point is clear: keep the agent on-device without sending sensitive context to the cloud.