Overview
Meta Muse Glimmer is a 30B dense multimodal model open-sourced by Meta Super Intelligence Lab on August 10, 2026, under the Apache 2.0 license. It is distilled from the flagship model Muse Spark 1.2 and is one of Meta's most permissively licensed open-weight models to date. Designed for local agent scenarios, it runs on a single consumer GPU, supports image, text, and video input, and tool calling, aiming to promote the distributed development of advanced AI. Its architecture combines alternating hybrid sliding window attention and full attention, GQA (reducing KV cache by 16x), 2D RoPE, and Pixel Shuffle to compress visual tokens, accompanied by a DFlash speculative decoding draft model for accelerated generation. The model leads in general agent and code scenarios, is available on OpenRouter, and developers have tested it running long contexts on a single RTX 4090.
Key Features
- Multimodal Input and Tool Calling: Supports image, text, and video input with up to 96 frames sampled from video, and features tool calling and object detection capabilities, suitable for complex agent tasks.
- Efficient Architecture Design: Employs alternating hybrid sliding window attention and full attention, combined with GQA to reduce KV cache by 16x, 2D RoPE for enhanced positional encoding, and Pixel Shuffle to compress visual tokens, improving long-context processing efficiency.
- Lightweight Deployment Friendly: The 4-bit quantized version is under 20GB, runs on a single consumer GPU or MacBook, compatible with NVIDIA, AMD, and Intel hardware, with GGUF quantization packages ready to use.
- Accelerated Inference Support: Comes with a DFlash speculative decoding draft model that significantly accelerates generation, especially beneficial in code and structured generation scenarios.
- Broad Ecosystem Compatibility: Adapted on day one for transformers, llama.cpp, vLLM, and HF Inference Endpoints, supporting LoRA and full fine-tuning, with fine-tuning requiring a minimum of a single 80GB H100.
- Permissive Open Source License: Released under the Apache 2.0 license, allowing free use, modification, and distribution, making it one of Meta's most permissively licensed open-weight models.
Use Cases
- Local agent development: Run on a single consumer GPU to build autonomous decision-making and tool-calling agents.
- Multimodal content understanding: Process mixed image, text, and video inputs for video analysis, image recognition, and object detection.
- Code generation and structured output: Leverage DFlash acceleration to improve code completion and structured data generation efficiency.
- Long-context tasks: Run long-context scenarios on RTX 4090, such as document summarization and multi-turn conversations.
- Research fine-tuning: Use TRL for LoRA or full fine-tuning to adapt to specific domain needs.
Pros
- Permissive open source license: Apache 2.0 facilitates commercial and academic use.
- Runs on a single GPU, lowering deployment barriers, suitable for individual developers and SMEs.
- Comprehensive multimodal capabilities: Supports video input and tool calling, rich in functionality.
- Complete inference acceleration support: Outstanding performance in code and structured generation.
- Broad ecosystem compatibility: Day-one support from mainstream frameworks, easy integration.
- Leading benchmark performance: MCP Atlas 75.5, SWE-Bench Pro 51.2, showing clear advantages in general agent and code scenarios.
Pricing
The model weights are free and open source; specific usage costs are subject to the official website.
Summary
Meta Muse Glimmer is an open-source multimodal model released under the Apache 2.0 license. It runs on a single GPU, supports image, text, and video input, and tool calling. It leads in general agent and code scenarios, has broad ecosystem compatibility, and is suitable for local deployment and fine-tuning.