Muse-Glimmer-30B
#44 v Open-source jazykové modelyUnsloth Ai · 30b · od 10. August 2026 · 2× · naposledy 13. 8. 2026
27
Momentum
Muse-Glimmer-30B is a 30-billion-parameter dense language model from Meta Superintelligence Labs featuring a dedicated perception encoder for multimodal input, distilled from the larger Muse Spark model. It was built for autonomous agentic workflows (tool use, multi-step reasoning, failure recovery) on consumer hardware and is released under the Apache 2.0 license. Unsloth provides quantized GGUF variants, including 2-bit versions for very limited VRAM. The model supports a context window of roughly 131,000 tokens and is offered through several third-party APIs as well as locally via llama.cpp/Unsloth.
Vývoj momenta
15.05.13.08.
Vlastnosti
| Key Benchmark (%) | SWE-Bench Pro: 51.2%; AIME 2026: 94.7% |
| Context Window (Tokens) | 131,072 tokens |
| License | Apache 2.0 |
| Multimodality | Yes – text and image understanding via a dedicated 1.8B perception encoder (ViT-G/14) |
| Platform | Hugging Face (unsloth/Meta), local via llama.cpp/Unsloth Desktop, plus APIs on OpenRouter, Together AI, ModelsLab, etc. |
| Price per 1M Tokens | $0.35 input / $1.50 output per million tokens (OpenRouter listing) |
| Release Date | August 10, 2026 |