Meta’s 30B-parameter Muse Glimmer runs offline on a single consumer GPU to execute local coding, function calling, and autonomous agent workflows.
On 10 August 2026, Meta Superintelligence Labs (MSL), led by Chief AI Officer Alexandr Wang, released Muse Glimmer under the Apache 2.0 licence on Hugging Face. Derived from the flagship Muse Spark teacher model, this release advances Meta’s open-source strategy against proprietary rivals like OpenAI and Anthropic, framing open weights as vital for American competitiveness and preventing regulatory capture.
Muse Glimmer is a 30-billion (30B) parameter dense multimodal model engineered for offline execution on consumer hardware. By applying 4-bit quantization, Meta compressed memory demands from 55 GB to 18–20 GB. This allows the model, KV cache, perception encoder, and speculative decoding drafter to run within a 24 GB or 32 GB VRAM envelope on a single consumer GPU, PC, or Mac. It features hardware optimizations for AMD, Arm, Dell, Intel, and Nvidia, alongside native runtime integration with llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, vLLM, and SGLang.
Trained via logit distillation, long-context agentic data, and reinforcement learning, Muse Glimmer supports text and image inputs while integrating with orchestration frameworks like OpenClaw. Optimized for agentic workloads, including coding, schedule management, file organization, function calling, and LLM-as-a-judge evaluations, it includes autonomous failure recovery to retry failed tool calls.
Following last month’s Muse Spark 1.1 release, Meta CEO Mark Zuckerberg confirmed plans to open-source the weights for the frontier-class Muse Spark 1.2 foundation model in the near future.
