Muse Glimmer 30B Released

Muse Glimmer has been released as a 30B dense multimodal open-weight model targeting local agent workflows. The weights are provided under the Apache 2.0 license, and it processes text and images together while supporting over 100 languages.

With 4-bit quantization, the language model memory usage is reduced to under 20GB, allowing the full configuration—including the KV cache, perception encoder, and reasoning drafter—to run in 24GB or 32GB memory environments. It also provides reasoning effort settings to adjust quality and speed.

Key features include the following:

• Support for agent tasks such as DeepSearch QA, MCP-Atlas, τ³-Bench, and SWE-Bench
• Precise function calling and multi-step reasoning in long-term workflows
• Error recovery that diagnoses and retries tool call failures or unexpected responses
• Integration with agent scaffolds such as OpenClaw
• Improved generation speed via DFlash-based speculative decoding

Weights are available on Hugging Face, with optimized integrations for Ollama, LM Studio, Unsloth, and torchtitan, as well as llama.cpp, MLX, and ExecuTorch, planned. Support for vLLM and SGLang serving, as well as usage through Together AI, Fireworks AI, and OpenRouter, is also planned.