Meta open-sources Muse Glimmer, a 30B agentic model that runs on one GPU

Meta Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter model aimed squarely at local, always-on agent workloads, with weights published on Hugging Face under a permissive Apache 2.0 license. The pitch is offline capability: the model targets a single consumer GPU in a 24–32 GB envelope, so function calling, coding, and LLM-as-judge tasks can run on a laptop or desktop without cloud infrastructure. To fit, Meta quantizes the weights to roughly 4-bit, cutting the footprint from over 55 GB to under 20 GB, and pairs it with a DFlash-based speculative-decoding drafter that proposes token blocks the main model verifies in parallel for faster generation at claimed-identical quality.

The model was built by distilling from a larger teacher (Muse Spark) via logit distillation, then mid-trained on longer-context, agent-heavy data and post-trained with a mix of supervised fine-tuning, on-policy distillation, and reinforcement learning. Meta emphasizes agentic behaviors over raw benchmark scores: end-to-end task completion on suites like SWE-Bench, τ-Bench, MCP-Atlas, and DeepSearch QA, reliable tool calling, multi-step planning, failure recovery, and multimodal input through a dedicated perception encoder. It also ships with controllable reasoning effort, 100+ language coverage, and compatibility with orchestration scaffolds. Meta positions it against Gemma4-31B and Qwen3.6-27B as strong for its size class.

The significance is less about topping leaderboards than about where inference happens. A capable, tool-using, multimodal agent that runs entirely on-device shifts personal-context workloads — schedules, files, screenshots, messages — off the network, which matters for privacy and offline use. Meta is leaning on an open ecosystem to drive adoption, with day-one or near-term support across Ollama, LM Studio, Unsloth, llama.cpp, MLX, ExecuTorch, vLLM, and SGLang, plus hardware tuning with AMD, Arm, Dell, Intel, and NVIDIA. The open-weight release continues Meta’s pattern of shipping usable models rather than API access, this time pushing that strategy into local agentic AI.