Overview
Meta Superintelligence Labs on August 10, 2026 released Muse Glimmer, <cite index="5-1,5-2">a 30-billion-parameter open agentic model whose weights are published under a permissive Apache 2.0 license.</cite> <cite index="5-3,5-4">The model is optimized for always-on local agent workflows and is small enough to run on a Mac or PC with a single consumer Graphics Processing Unit (GPU), enabling use cases that range from local agents and function calling to local coding and Large Language Model (LLM)-as-a-judge evaluation.</cite>
Architecture and Quantization
<cite index="15-10,15-11">At full precision, a 30-billion-parameter model would require over 55 GB of memory — far more than any consumer GPU offers. Meta uses quantization techniques to compress the model's weights to approximately 4-bit precision, shrinking the language model to under 20 GB.</cite> <cite index="15-12,15-13">This leaves enough headroom for the model's working memory (its KV cache), the perception encoder for image understanding, and the speculative-decoding drafter to run simultaneously within a 24 GB or 32 GB envelope, with Meta validating that this compression introduces minimal to no degradation on agentic tasks.</cite>
<cite index="6-12">A lightweight DFlash-based companion model proposes blocks of tokens that the main model then verifies in parallel</cite>, enabling practical inference speeds on local hardware. <cite index="1-10">Early performance testing shows Muse Glimmer 30B reaching up to 24 tokens per second on an AMD Ryzen AI Max+ 395 processor.</cite>
Training Pipeline
<cite index="9-8">Muse Glimmer is distilled from Muse Spark, the model that Meta Superintelligence Labs presented in April of this year.</cite> <cite index="9-9">Training took place in three phases: logit distillation on the output of the larger model, mid-training with longer context and agentic data, and finally, supervised fine-tuning combined with reinforcement learning.</cite> <cite index="16-4">Meta evaluated Muse Glimmer under the standards of its Advanced AI Scaling Framework and assessed it for open-weight release across the relevant categories.</cite>
Agentic Capabilities
<cite index="4-3">Meta's Chief AI Officer Alexandr Wang stated that the model supports agentic use through planning, tool calls, result checking, and failure recovery.</cite> <cite index="1-12,1-13">Muse Glimmer 30B is designed for longer-arc workflows, including complex workflows that span multiple steps and sessions; it can manage memory, recover from failures, and continue across restarts, helping developers move beyond one-shot interactions toward agents that can take ownership of meaningful work.</cite> <cite index="14-5">The model supports over 100 languages, positioning it for global deployment rather than the English-first approach that characterized earlier lightweight models.</cite>
Benchmarks and Comparisons
<cite index="16-13">The model was compared with Gemma4-31B and Qwen3.6-27B and, according to Meta, performs strongly for its size class on several widely used LLM benchmarks.</cite> <cite index="16-6,16-7">Evaluation benchmarks include DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which measure the model's ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish.</cite>
Availability and Ecosystem
<cite index="6-4">The model is available on Hugging Face under the Apache 2.0 license.</cite> <cite index="11-13">Optimized support for llama.cpp, MLX, and ExecuTorch is due in the coming days, alongside planned access through Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter.</cite> <cite index="11-14">Meta is also working with AMD, Arm, Dell, Intel, and NVIDIA on device-level optimization, extending its open AI research into local agent systems.</cite>
<cite index="4-4">Meta's Chief AI Officer also indicated that open weights for Muse Spark 1.2 will follow soon.</cite>