Muse Glimmer is Meta’s newest open model, and it aims at your desktop. Meta introduced the 30B agentic model on Monday, built to run always-on agents on a single consumer GPU. In addition, the weights are free on Hugging Face under the Apache 2.0 license.
Built for agents, sized for one GPU
Muse Glimmer descends from Muse Spark, Meta’s closed flagship. Logit distillation taught the smaller model to mimic its teacher on agent tasks. As a result, it handles multi-turn requests, precise tool calls and long multi-step plans. It also recovers from failed tool calls, and it reads interleaved images through a dedicated perception encoder.
The memory math is the headline. In practice, full precision needs more than 55 GB, beyond consumer cards. However, a 4-bit quantized build keeps the language model under 20 GB. Hugging Face lists a K-Quant-Dynamic cut that fits 32 GB cards. A 17 GB flavor suits 24 GB boards, leaving room for the KV cache, encoder and drafter. Specifically, Meta tested MacBook M4-Max and M5-Max machines plus an NVIDIA RTX 5090.
Moreover, the model trained on data from more than 100 languages. It also offers controllable effort, so agents can trade reasoning depth for speed. Scaffolds get first-class treatment too. The model slots into OpenClaw and similar orchestrators, and developer docs live at dev.meta.ai. The license allows commercial and research use.
Local execution is the pitch beyond price. In other words, agents that read your files and calendars never leave the machine. For privacy-minded users, that beats any cloud API.
How Meta shrank Spark
Training followed three acts. First, pre-training distilled Muse Spark outputs. Next, mid-training added longer-context, agent-heavy data with richer reasoning traces. Post-training then mixed supervised fine-tuning, on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains.
For speed, Meta ships quantized weights plus a lightweight DFlash speculative-decoding drafter. Safety work followed Meta’s Advanced AI Scaling Framework. Still, one asterisk matters: the weights are open, yet the Spark teacher stays closed. That is why Engadget put “open source” in quotes.
Benchmarks, partners and the open-source race
Meta evaluated Muse Glimmer on DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench, against Gemma4-31B and Qwen3.6-27B. Yet the blog publishes no scores; a separate methodology report carries the numbers. So readers must take “strong success rates” on faith for now.
The partner bench is long. Optimized builds for llama.cpp, MLX and ExecuTorch land in the coming days. Ollama, LM Studio and Unsloth cover local runs, while vLLM and SGLang handle serving. Together AI, Fireworks and OpenRouter will host it, and AMD, Arm, Dell, Intel and NVIDIA signed on for hardware tuning.
The release lands in a crowded field. According to Engadget, Meta’s Muse Spark trails rivals from OpenAI and Anthropic, while Chinese models like DeepSeek already run local; Tech My Money covered DeepSeek’s permanent V4 price cut in May. Instead, Muse Glimmer answers with openness and local control. Mark Zuckerberg framed the mission in an essay: “Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.” Meta’s own Muse family keeps shipping too, most recently Muse Code, a terminal agent on Spark 1.2.
