Meta has released Muse Glimmer, a new open-source multimodal language model designed for local agentic use cases. Distilled from a larger Muse model to 30B parameters and released under the Apache 2.0 license, it targets privacy-aware applications like coding, document analysis, and personal assistants.

Muse Glimmer pairs a 2B ViT-style Perception Encoder for vision with a 28B text decoder. The text decoder uses a hybrid attention pattern alternating three sliding window layers (RoPE, 2,048 tokens) with one full attention layer (NoPE), repeated 13 times for 52 total layers. Gated Grouped-Query Attention shares one key-value head across 16 query heads, reducing KV-cache memory by 16x. Q-K normalization with an extra query scale factor stabilizes attention logits. The Perception Encoder handles both images and videos, processing 2 frames per second with a cap of 96 frames per video. An optional speculative decoding drafter (DFlash) can accelerate structured generation like code.

Benchmarks show competitive results on agentic, coding, and multimodal tasks against models like Gemma4-31B (thinking mode) and Qwen3.6-27B (thinking mode). Notably, Muse Glimmer scores 75.5 on MCP Atlas (vs. 54.2 and 62.5), 76.0 on SWE-Bench Verified (vs. 66.6 and 77.2), and 94.7 on AIME 2026 (vs. 89.2 and 94.1). Safety benchmarks show a CI Memories violation rate of 26.4% with 64.8% coverage, and a Siren AgentDojo attack success rate of 28.4% with 94.2% utility.

The model ships with day-0 support in transformers, llama.cpp, vLLM (with transformers backend), and Inference Endpoints. Hugging Face provides example code for text-only inference, multimodal prompting, tool calling, object detection, and video inference. Fine-tuning support via TRL includes SFT and Async GRPO, with practical minimum hardware recommendations ranging from a single 80GB H100 for LoRA to eight 80GB H100s for full fine-tuning. Several creative demos show Muse Glimmer autonomously quantizing itself, deploying itself to Inference Endpoints, optimizing its own inference stack, and researching the Hugging Face Hub as an agent.