Summary

On August 10, 2026 Meta Superintelligence Labs released Muse Glimmer under the Apache 2.0 license on Hugging Face, a 30-billion-parameter dense multimodal model engineered for offline agent workflows on consumer hardware. With 4-bit quantization Meta cut memory needs from about 55 GB to 18 to 20 GB so the model, KV cache, perception encoder, and speculative-decoding drafter fit within a 24 GB or 32 GB VRAM budget on a single GPU, PC, or Mac, and it runs via llama.cpp and Ollama.

What changed

Meta published Muse Glimmer, a 30B open-weights (Apache 2.0) multimodal model distilled from its larger Muse Spark series and optimized for always-on local agents: coding, function calling, schedule and file management, and multi-step reasoning with failure recovery, downloadable on Hugging Face and runnable offline on a single consumer GPU.

Why it matters

A capable 30B agentic model that runs offline on everyday hardware pushes autonomous agents toward local and edge execution, cutting inference cost and data-exposure concerns and challenging the assumption that useful agents require hosted frontier APIs. Apache 2.0 licensing makes it a base others can build and ship on, strengthening the open-model side of the agent stack.

Evidence excerpt

August 10, 2026: Meta Superintelligence Labs released Muse Glimmer under Apache 2.0 on Hugging Face, a 30B dense multimodal model for offline local agents; 4-bit quantization compresses memory from ~55 GB to 18-20 GB to fit a 24 GB or 32 GB VRAM envelope on a single consumer GPU, PC, or Mac, deployable via llama.cpp and Ollama, distilled from Meta's Muse Spark series.

Sources