Meta's Muse Glimmer Is a 30-Billion-Parameter AI You Can Run on Your Own Machine
The new open-source model sees images, understands video, and can act as a personal assistant without sending your data to a cloud server.
Key points
- Meta released Muse Glimmer today under the Apache 2.0 open-source licence, meaning anyone can use or modify it for free.
- The model has 30 billion parameters, a rough measure of its complexity, making it smaller and cheaper to run than many frontier AI systems.
- Muse Glimmer processes text, images, and video inside a single model and can call external tools like a weather service on its own.
- Hugging Face shipped support on day one across five major software frameworks.
- A built-in speed booster called a speculative decoding drafter is especially good at generating structured output like code.
Meta just released a new AI model, and the headline fact is where it runs: on your own computer, not a company's server.
Muse Glimmer is a 30-billion-parameter multimodal model, distilled from a larger Meta model called Muse. Think of distillation as producing a concise student edition from a sprawling textbook. The result is compact enough to run locally while still handling complex tasks. We first covered Muse Glimmer on 10 August 2026; this release is the open weights drop the community has been waiting for.
What can it actually do?
Muse Glimmer can answer questions about photos, analyse documents, write code, and act as a personal assistant that takes actions on your behalf. Show it a picture of a city and ask what to pack, and it can recognise the location then automatically query a weather service for a real answer.
Video is supported too. Footage is processed at up to 96 frames per clip at two frames per second, with each frame stamped with its timestamp so the model understands when things happen, not just what is shown.
It's also what engineers call agentic: it can break a goal into steps and carry them out without someone directing each one by hand. Less like a search engine, more like a capable intern who can read things and act on them.
Why does running it locally matter?
Your documents, photos, and conversations never leave your machine. That matters for anyone handling sensitive information: medical notes, legal contracts, private code repositories.
It also cuts cost. Cloud AI services charge per request; a locally-run model costs only the electricity to run it. The Apache 2.0 licence lets businesses build commercial products on top of Muse Glimmer without paying royalties, which is the part that will matter most to developers who want to ship something fast.
How fast is it?
Inside the model sits an optional speed module called a speculative decoding drafter, built on a system called DFlash. Speculative decoding works by having a smaller, faster model guess the next several words while the main model checks those guesses in bulk, cutting response time. The trade-off is slightly more memory use. Meta says the drafter works especially well when generating structured content like code.
Hugging Face confirmed day-one support across transformers, llama.cpp (a tool designed to run large models on ordinary consumer hardware), vLLM, and two other libraries. The same code runs unchanged on NVIDIA, AMD, and Intel graphics cards, the specialised chips AI models rely on for number-crunching.
What happens next?
Meta is targeting developers and privacy-conscious users who want powerful AI without a subscription or a data-sharing agreement. The open-weights local AI space is getting crowded fast. Our story on Liquid AI's 2.6B agent model showed how capable small local models have become; Muse Glimmer's 30B scale is a significant step up in ambition, and the real test is whether consumer hardware can actually sustain it at useful speeds.
The full model collection is available on Hugging Face.
