Muse Glimmer (high)

Muse Glimmer is a 30-billion-parameter multimodal language model developed by Meta Superintelligence Labs and released under the Apache 2.0 license. Purpose-built for autonomous agentic tasks and local deployment, the model is designed to handle long-horizon workflows including multi-step reasoning, sequential tool use, and failure recovery. It represents a shift in Meta's release strategy, providing fully open weights and a permissive license for the first time since the Llama series.

The model's architecture is a dense causal transformer distilled from Meta's proprietary Muse Spark flagship. It incorporates a ~1.8B-parameter vision encoder, allowing it to process interleaved text and images such as screenshots, documents, and charts. Muse Glimmer utilizes a specialized hybrid-attention mechanism combining Grouped-Query Attention (GQA) and Sliding Window Attention (SWA) with an extreme query-to-KV head ratio, significantly reducing the KV-cache footprint and enabling efficient inference on consumer-grade hardware.

Reasoning and Capabilities

A core feature of Muse Glimmer is its support for Controllable Reasoning Effort, which allows developers to select from four distinct reasoning strengths: low, medium, high, and xhigh. The high reasoning setting is optimized for complex tasks that require extensive chain-of-thought processing and sustained planning across extended multi-turn interactions. Benchmarks indicate that in this mode, the model effectively matches the reasoning performance of much larger proprietary systems while maintaining a local-friendly footprint.

Muse Glimmer is specifically tuned for agentic reliability, with training focused on the ability to diagnose failed tool calls and retry operations rather than halting execution. It features a default context window of 131,072 tokens, which can be extended for tasks involving large codebases or long document histories. The model was trained on data spanning over 100 languages and has a knowledge cutoff date of January 4, 2026.