Meta releases Muse Glimmer, a 30B open-weight agentic model that runs on a single consumer GPU

Archive entry: this covers an event from August 10, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On August 10, 2026, Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter open agentic model under the Apache 2.0 license, as described in Meta's research blog post. Meta said the model ran on a Mac or a PC with a single consumer GPU.

What changed

• Weights and docs: the weights were on Hugging Face as meta-models/Muse-Glimmer-30B, with documentation at dev.meta.ai/docs/muse-glimmer.
• Size: full-precision weights exceeded 55 GB. Quantization to about 4 bits brought the model under 20 GB, which fit 24 or 32 GB machines together with the KV cache, perception encoder and drafter.
• Capabilities: the model accepted image input, supported 100+ languages, offered controllable reasoning effort and shipped with a DFlash speculative-decoding drafter.
• Training: logit distillation from Muse Spark, followed by SFT, on-policy distillation and RL.
• Comparisons: Meta compared the model with Gemma4-31B and Qwen3.6-27B on DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench, and measured speed with a K-Quant 17 GB build on M4 Max, M5 Max and RTX 5090 hardware.

What it meant for people building with AI

Meta shipped an Apache 2.0 agentic model built to run locally on developer hardware: the roughly 4-bit build was under 20 GB and fit 24 or 32 GB machines, and Meta said a single consumer GPU was enough.

Launch partners

Launch partners included Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, Together, Fireworks and OpenRouter. Hardware partners were AMD, Arm, Dell, Intel and NVIDIA.

Sources

• Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device (Meta AI Research, August 10, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.