Key takeaways

• Muse Glimmer 30B is a multimodal model from Meta, optimized for autonomous agents on consumer hardware. It supports interleaved text and image inputs with a 131K-token context window and is designed for long-horizon agentic tasks including reliable tool use and failure recovery.
• It utilizes a 29.6B-parameter dense causal transformer with a dedicated 1.8B-parameter vision encoder, indicating a Transformer-based architecture for multimodal processing. It is open-weight and released under the Apache 2.0 license.
• The model excels in problems requiring algebraic manipulation, geometric reasoning, and number theory, achieving 86.67% accuracy on the AIME 2026 benchmark. It also performs well in qualitative and general knowledge tasks, particularly in visually-rich, descriptive domains (e.g., 88% accuracy in Art on MMMU benchmark).
• Key limitations include struggles with complex combinatorial problems, calculation errors in quantitative reasoning, and misinterpretation of complex diagrams in STEM and Finance. It often fails when precise calculations and multi-step logic are required.
• The model achieved 86.67% accuracy on the AIME 2026 benchmark, showing strengths in various mathematical domains but struggling with complex combinatorial conditions. On the MMMU benchmark, it showed strong performance in domains like Art (88% accuracy) and Agriculture, but struggled significantly with quantitative reasoning in Finance, Accounting, and Engineering.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

Transformer

Modalities

text+image->text

Parameters

~30B

Open weights

Yes

Licence

Apache 2.0

Regions

USA

Vendor

Meta

Provider model key

meta/muse-glimmer-30b