Muse Glimmer 30B added to BenchGecko · Meta's open weights agent model under Apache 2.0
Meta published the Muse Glimmer 30B paper on 2026-10-19. The weights have been on Hugging Face since 2026-08-09 under Apache 2.0. BenchGecko lists a 131,072 token context and hosted rates from $0.30 input and $1.20 output per M tokens.
What launched
Meta published a research paper on Muse Glimmer 30B on 2026-10-19. The model itself shipped earlier: Meta Superintelligence Labs put the weights on Hugging Face as meta-models/Muse-Glimmer-30B on 2026-08-09, and the model card gives August 2026 as the release date. BenchGecko records 2026-08-09 as the release date. The model was added to BenchGecko today under Meta.
Muse Glimmer is a dense model distilled from Muse Spark and built for agentic tasks that run locally on a single GPU or Mac. It reads text and images and answers in text. Meta says it was trained on data from more than 100 languages.
Specs
From the Hugging Face model card:
- Parameters · about 29.6B, including a perception encoder of about 1.8B
- Architecture · dense causal transformer, 52 layers, sliding window attention over 2,048 tokens on three of every four layers
- Context · 131,072 tokens
- Inputs · text and image
- Output · text
- Reasoning strength · low, medium, high and xhigh
- Knowledge cutoff · 2026-01-04
- Licence · Apache 2.0
Meta ships BF16 weights, two 4-bit builds and a DFlash speculative decoding drafter. The 4-bit builds bring the language model under 20 GB and target 24 GB and 32 GB of VRAM, with an average degradation of 1.0% and 0.2% across 15 benchmarks. Full precision targets 64 GB.
Pricing on BenchGecko
Meta distributes Muse Glimmer as downloadable weights. OpenRouter lists two hosted endpoints, both serving the full 131,072 token context:
- DeepInfra · bf16 · $0.30 per M input tokens and $1.20 per M output tokens, cached input $0.04
- Together · $0.35 input and $1.50 output, cached input $0.04
BenchGecko stores the DeepInfra rates on the model row. For scale, Gemma 4 31B lists at $0.09 input and $0.34 output on BenchGecko, and Qwen3.6 27B at $0.25 and $2.00. Both have a 262,144 token context.
Compare rates · BenchGecko pricing.
Vendor-reported benchmarks
Meta reports these results on the model card for Muse Glimmer at high reasoning strength, next to Gemma4-31B and Qwen3.6-27B in thinking mode. BenchGecko records them as vendor claims. No independent scores or Gecko Tests are attached to Muse Glimmer yet.
- MCP Atlas (Public) · Muse Glimmer 75.5 · Gemma 4 31B 54.2 · Qwen3.6 27B 62.5
- DeepSearch QA · Muse Glimmer 74.6 · Gemma 4 31B 61.7 · Qwen3.6 27B 71.1
- SWE-Bench Pro · Muse Glimmer 51.2 · Gemma 4 31B 36.9 · Qwen3.6 27B 50.2
- SWE-Bench Verified · Muse Glimmer 76.0 · Gemma 4 31B 66.6 · Qwen3.6 27B 77.2
- TerminalBench 2.1 · Muse Glimmer 51.7 · Gemma 4 31B 43.4 · Qwen3.6 27B 60.7
- OSWorld-Verified · Muse Glimmer 65.9 · Gemma 4 31B 58.5 · Qwen3.6 27B 75.6
- AIME 2026 · Muse Glimmer 94.7 · Gemma 4 31B 89.2 · Qwen3.6 27B 94.1
- GPQA Diamond · Muse Glimmer 83.5 · Gemma 4 31B 85.7 · Qwen3.6 27B 84.2
- MMMU Pro · Muse Glimmer 74 · Gemma 4 31B 73 · Qwen3.6 27B 75
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.