Meta has released 'Muse Glimmer,' an open-source model that runs locally, and claims it performs better than Google's Gemma 4 31B on the 30B.



Meta released its AI model ' Muse Glimmer, ' with 29.6 billion parameters, as an open model on August 10, 2026. It is being touted as being more powerful than comparable models from other companies.

Muse Glimmer | Meta

https://developer.meta.com/ai/models/muse-glimmer/

Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device | Meta AI Research
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model

Muse Glimmer is a Dense model with 29.6 billion parameters and supports text and image input. Below is a table comparing the benchmark scores of 'Muse Glimmer 30B,' 'Gemma 4 31B,' and 'Qwen 3.6-27B.' Muse Glimmer outperforms Gemma 4 31B in many tests.



Third-party organization

Artificial Analysis has also published its test results for Muse Glimmer. In the Artificial Analysis Intelligence Index, an index that aggregates multiple benchmark scores, it surpassed the closed model Claude Haiku 4.5 and recorded a score approaching that of Gemini 3.5 Flash-Lite. This is a significant leap compared to the Llama 4 series, which was released in 2025.



The graph below shows the model size on the horizontal axis and the Artificial Analysis Intelligence Index score on the vertical axis. It can be seen that Muse Glimmer is smaller in size compared to models with comparable performance.



Muse Glimmer supports

DFlash , a speculative decoding technique that generates drafts using a lightweight diffusion model, enabling high-speed processing even locally. With an NVIDIA GeForce RTX 5090, using DFlash improves decoding speed by 3.1 times, allowing for decoding of 233 tokens per second.



Muse Glimmer is available at the following link and can load the entire waitlist onto a GPU with 24GB of VRAM and run locally. The license is the Apache License 2.0.

Muse Glimmer - a meta-models Collection
https://huggingface.co/collections/meta-models/muse-glimmer



AI execution tools such as LM Studio and Ollama have already completed support for Muse Glimmer. Additionally, quantization models created by volunteers are compiled at the following link.

Quantized Models for meta-models/Muse-Glimmer-30B – Hugging Face
https://huggingface.co/models?other=base_model:quantized:meta-models/Muse-Glimmer-30B



The Muse Glimmer documentation is available at the following link.

Overview | Model API
https://dev.meta.ai/docs/muse-glimmer/



Meta has also announced that they will be making Muse Spark 1.2 an open model in the near future.




Related Posts:

in AI, Posted by log1o_hf