Muse Glimmer
Meta · Released Aug 2026
Muse Glimmer is Meta's 30B open-weight agentic model, distilled from the closed Muse Spark frontier model and released under Apache 2.0. With 4-bit quantization it runs multi-step tool calling, multimodal reasoning, and failure recovery on a single 24GB consumer GPU without cloud dependencies. It leads its size class on agentic orchestration but trails Qwen3.6-27B on terminal and OS automation — reach for it if you need a capable local agent that stays offline, not if you need independent reproducibility or low-level OS control.
Is it right for you?
Good for
- Local agentic workflows — runs on a single 24GB consumer GPU at ~20GB quantized, no cloud or internet required
- Tool use and function calling — reliable schema-based tool invocation across long multi-turn agent sessions
- Multimodal reasoning — includes a 1.8B ViT-G/14 perception encoder for screenshots, charts, and documents
Not good for
- Terminal/OS automation — trails Qwen3.6-27B on TerminalBench 2.1 (51.7 vs 60.7) and OSWorld-Verified (65.9 vs 75.6)
- Reproducibility-critical use — Apache 2.0 weights only; no training data or code released, so the model cannot be independently reproduced
How it performs by task
Agentic task completion
Leads Gemma4-31B and Qwen3.6-27B on MCP Atlas (75.5) and DeepSearch QA (74.6); strong multi-step agent reasoning
Code generation
SWE-Bench Verified 76.0, SWE-Bench Pro 51.2 — competitive with Qwen3.6-27B on verified, ahead on Pro
Math reasoning
AIME 2026 score of 94.7, matching Qwen3.6-27B (94.1) and exceeding Gemma4-31B (89.2)
Terminal/CLI automation
TerminalBench 2.1 score of 51.7 — meaningfully behind Qwen3.6-27B (60.7); not the best fit for shell-heavy agent work
Pricing
Input
Free (open weights)
Output
Free (open weights)
Context
131K tokens
Benchmarks
No verdict changes yet
The clock starts day one — changes land here as our verdict evolves.
Sources
Verification log
No verification checks yet
We haven't logged a verification check for this entry. Once a check runs, its history shows here.