Apache 2.0 Licensed 30B Muse Glimmer Drives Meta's Local Agents

On August 10, 2026, Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter multimodal model aimed at local agentic workflows. The launch of Muse Glimmer arrived alongside CEO Mark Zuckerberg’s manifesto on personal superintelligence. For developers, the real news is the licensing. Meta dropped the restrictive Llama License in favor of Apache 2.0.

Technical Specifications

Muse Glimmer is a dense causal transformer with 29.6 billion total parameters. It avoids a mixture-of-experts approach, instead utilizing a dedicated 1.8-billion-parameter perception encoder for multimodal image inputs. The model features 52 layers, a hidden size of 6,656, and a 131,072-token context window.

Running the model requires managing its weights based on the target hardware. In BF16, the model consumes approximately 60 GB of memory. Using 4-bit quantization reduces the footprint to 18 GB, allowing it to fit comfortably on modern consumer GPUs.

Hardware Benchmarks and Performance

Muse Glimmer demonstrates significant speed when deployed across both consumer and enterprise hardware.

Hardware	Performance
AMD Radeon AI PRO R9700	53 tokens/sec
NVIDIA Blackwell Ultra	20,000 tokens/sec

The model scored 35 on the Artificial Analysis Intelligence Index. This places it 21 points higher than Meta’s previous open release, Llama 4 Maverick. The reasoning performance effectively matches the 1-trillion-parameter Kimi K2.5 despite being 33 times smaller.

Governance and Licensing

The shift to Apache 2.0 signals a change in Meta’s distribution strategy. Analysts point out this positions Muse Glimmer as a mechanism for local AI deployment on hardware Meta does not control, targeting market share currently held by Microsoft and Google.

Meta also announced its independent board now holds authority over model-release safety criteria. To address privacy concerns regarding personal agents, the company promised a fully private mode where Meta cannot access user data.

If you are building agentic systems for local execution, Muse Glimmer offers frontier-level reasoning with a fully permissive license. The upcoming open-weight release of the flagship Muse Spark 1.2 will further expand the options for on-device computing.