Meta releases Muse Glimmer, an open-source 30B multimodal agentic model for local use

Meta has released Muse Glimmer, a 30B-parameter multimodal model designed for local agentic tasks, available under Apache 2.0 and optimized for single-GPU deployment.

This article was drafted with AI assistance from multiple sources and was reviewed and approved by a human editor before publication.

Meta released Muse Glimmer on August 10, 2026, an open-source multimodal model with 30 billion parameters, targeting local agentic use cases. The model is distilled from Meta's larger Muse model and is available under the Apache 2.0 license, with weights free on Hugging Face Hub.

Muse Glimmer is a dense transformer model that incorporates a 2B-parameter Vision Transformer-like Perception Encoder for processing images and videos. The encoder patchifies images into 2 frames by 3 channels by 14 by 14 patches, and its vision tower consists of 50 layers with GELU MLPs. Attention is structured with three window attention layers followed by one full attention layer, and 2D Rotary Position Embedding (RoPE) is applied to queries and keys. A pixel shuffle mechanism concatenates 2x2 groups of neighboring spatial tokens, reducing image token count by four times. For video, the processor targets 2 frames per second and caps clips at 96 frames.

The model supports text-only inference, image and text input, video, multimodal tool calling, and object detection. It is based on Meta's closed Spark 1.2 model and was trained on data from over 100 languages. Designed to balance capability against local hardware constraints, Muse Glimmer can run on a single GPU, requiring a Mac or PC with a dedicated GPU. It is available through Hugging Face, Ollama, LM Studio, and Unsloth, with day-0 support in transformers, llama.cpp, vLLM, and Inference Endpoints.

Muse Glimmer includes agentic capabilities for multi-step tasks, tool use, and failure recovery. An optional speculative decoding drafter, implemented on DFlash, can accelerate inference; it uses a block size of 16 (one anchor plus 15 proposed tokens). Meta's tests show that with speculative decoding, the model outputs over 57 tokens per second, compared to about 26 tokens per second for comparable models without it.

Meta benchmarks indicate that Muse Glimmer outperforms Gemma 4 and Qwen3.6 in general agentic skills, agentic coding, and reasoning. However, such benchmarks from the developer should be viewed with skepticism, as they may be biased.

In a company statement, Meta said, "We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware." Mark Zuckerberg added, "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it."