Muse-Glimmer-30B
#33 in Open-Source LLMsUnknown · 30b · since 10. August 2026 · 3× · last seen Aug 18, 2026
34
Momentum
Muse Glimmer 30B is an open-weight, multimodal language model from Meta Superintelligence Labs with roughly 30 billion parameters (dense architecture, including a ~1.8B vision encoder). It was distilled from the larger closed-source Muse Spark model and is purpose-built for local agentic workloads (tool use, multi-step planning, failure recovery) on consumer hardware such as Macs or PCs with capable GPUs. The model accepts text and images, supports over 100 languages, and is released under the permissive Apache 2.0 license. It is hosted on platforms including Hugging Face, NVIDIA NIM, Together AI, and OpenRouter, with a context window of 128K to 131,072 tokens.
Momentum trend
20.05.18.08.
Features
| Key Benchmark (%) | 75.5% on MCP-Atlas (Public); 74.6% on DeepSearch QA; 94.7% on AIME 2026 |
| Context Window (Tokens) | 131,072 tokens (128K+) |
| License | Apache 2.0 |
| Multimodality | Text and image input, text output (dedicated perception encoder for images) |
| Platform | Hugging Face, NVIDIA NIM/build.nvidia.com, Together AI, OpenRouter, llama.cpp, vLLM, Unsloth, Jetson AI Lab |
| Price per 1M Tokens | approx. $0.32 (input) / $1.35 (output) per 1M tokens (median across hosting providers; model itself free/Apache 2.0) |
| Release Date | August 10, 2026 |