Meta has released Muse Glimmer, a 30-billion-parameter multimodal model distilled from its larger Muse model and published under the Apache 2.0 license, according to a Hugging Face blog post dated August 10, 2026. Hugging Face describes the model as designed for local agentic use cases such as coding, document analysis, and personal assistants, and says it shipped with day-0 support in transformers, llama.cpp, vLLM, and Inference Endpoints. The model is a dense 30B architecture that uses a single 2B ViT-like Perception Encoder to process both images and video, with the video processor targeting 2 frames per second and capping clips at 96 frames. An optional speculative decoding drafter is included to speed up generation, which Hugging Face reports is well suited to structured content such as code. Developer Simon Willison, writing on August 10, 2026, tested the model locally using LM Studio's 18.16 GB version and ran it against a codebase with his coding-agent plugin. Willison noted the Apache 2.0 license is 'a step up from the janky Llama licenses of old,' and said the 30B size fits on machines with 32 GB of RAM or more. He also tested its image-description capabilities.
- 30B-parameter multimodal model distilled from Meta's Muse, released under Apache 2.0 on August 10, 2026
- Day-0 support in transformers, llama.cpp, vLLM, and Inference Endpoints; LM Studio version is 18.16 GB
- Uses a 2B Perception Encoder for images and video; runs comfortably on machines with 32 GB RAM or more
What it means for you
Meta has released a capable AI model you can run on your own computer for free, with a genuinely permissive license (Apache 2.0) that lets businesses use it commercially without the licensing headaches its older models carried. It handles text, images, and video, and is sized to run on a decent laptop or desktop with 32 GB of memory. For most people, this changes nothing today — but it's a real option for anyone who needs AI that keeps data on their own machine.
Try this
If you have a machine with 32 GB of RAM or more and want to test private, offline AI, download LM Studio and load the Muse Glimmer model to see how it handles a document or coding task you'd otherwise send to a cloud service.
Who should care
Developers and technically comfortable small businesses that want to run AI locally — for privacy, cost control, or handling sensitive documents that shouldn't leave the building.
Skip this if
You're happy using cloud tools like ChatGPT or Claude, don't have the hardware to run models locally, or have no one on hand who is comfortable installing developer software.
Sources: Hugging Face, Simon Willison — read the original