Muse-Glimmer-30B: 30B Multimodal Model Achieves 76% on SWE-bench

Muse Glimmer-30B is a 30B parameter multimodal language model that simultaneously understands text, images, and video. Capable of processing image patches and video tokens in addition to text inputs, it performs complex tasks mixing visual information and text instructions within a single model. Built on the Transformers library, it supports immediate inference within the Hugging Face ecosystem.

The model's core strength lies in its structured tool calling and separated reasoning processes. Through the ATEM (Advanced Tool-calling Environment Model) protocol, tool calls are separated into a distinct channel, while internal reasoning is independently logged to the 'to=self' channel. This ensures strict argument validation during tool calls and helps reduce errors in complex workflows. It has demonstrated code modification capabilities by scoring 76% on SWE-bench Verified and 51.2% on SWE-bench Pro.

Unlike existing single-channel tool calling methods, it adopts a structure that explicitly distinguishes recipients. Message flows between the system, user, tools, and assistant are clearly separated, making it easier to maintain context in environments where multi-turn conversations and tool integrations are mixed. With scores of 83.5% on the GPQA Diamond benchmark and 94.7% on AIME 2026, it shows strength in high-difficulty reasoning and mathematical problem solving.

Real-time inference is supported by Together AI and Fireworks AI. Notably, Together AI offers a processing speed of 96.7 tokens per second with structured output and tool calling features enabled. It is suitable for automatic code correction, multimodal document analysis, and agent development requiring complex tool integration. Deployable on-premises at the 30B scale, it has recorded over 430,000 downloads, indicating community validation.