On August 10, 2026, Meta released Muse Glimmer — an open model with 30 billion parameters that can run on a single consumer GPU without connecting to the cloud. We break down what this model is and how it differs from the closed Muse Spark.
What Is Muse Glimmer
Muse Glimmer is a model obtained through distillation from the closed flagship Muse Spark, which Meta introduced back in April 2026. Unlike Muse Spark, access to which is possible only through the company’s API, Muse Glimmer’s weights have been published under the Apache 2.0 license — meaning they can be freely downloaded, modified and run independently.
The model has about 29.6 billion parameters in a dense transformer architecture, supplemented by a separate visual encoder with 1.8 billion parameters, which allows the model to read images, screenshots, charts and documents. The context window is 131,072 tokens, with support for more than 100 languages.
Main Feature: Runs on a Regular Computer
At full precision, a 30-billion-parameter model usually requires about 55 GB of memory — a server-grade hardware level. Meta engineers compressed the model to less than 20 GB using quantization, allowing it to run on a single consumer graphics card — for example, an Nvidia RTX 3090 or AMD 9700 — or on a Mac with an M4/M5 Max-series chip, without any network access.
A separate technical detail is speculative decoding, which reduces response latency, making the model fast enough to work in real agentic scenarios.
Benchmarks: Strong in Some Areas, Weaker in Others
According to Meta’s internal tests, Muse Glimmer outperforms similarly sized models — Gemma4-31B and Qwen3.6-27B — on roughly half of two dozen popular benchmarks, especially in online research, code generation and scientific chart analysis tasks.
At the same time, the model trails competitors in computer-use tasks and terminal commands — meaning it is not a universal leader, but a model with clear strengths and weaknesses typical for its size class.
What Else Meta Announced
Along with the release of Muse Glimmer, Meta announced that in the coming weeks it will open the full weights of Muse Spark 1.2 itself — the company’s most powerful closed model. This is Meta’s first open release in more than a year after the previous generation of Llama models.
In Brief: The Key Points
- August 10, 2026 — Muse Glimmer release, 30B parameters, Apache 2.0 license
- Compressed to less than 20 GB — runs on a single consumer graphics card or a Mac with an M4/M5 Max chip
- Context window — 131,072 tokens, support for 100+ languages
- Outperforms Gemma4-31B and Qwen3.6-27B on roughly half of the benchmarks, but trails in computer-use/terminal tasks
- Meta also plans to open the weights of Muse Spark 1.2 in the coming weeks
FAQ
What is Meta Muse Glimmer? An open AI model with 30 billion parameters, obtained by distillation from the closed Muse Spark model and published under the Apache 2.0 license.
Can Muse Glimmer run on a regular computer? Yes, after quantization the model takes up less than 20 GB and runs on a single consumer graphics card or a Mac with an M4/M5 Max-series chip, without connecting to the cloud.
How is Muse Glimmer different from Muse Spark? Muse Spark is Meta’s closed top-tier model, available only through an API. Muse Glimmer is a smaller, open version obtained through distillation, which can be downloaded and run independently.
How good is Muse Glimmer compared with competitors? It outperforms Gemma4-31B and Qwen3.6-27B on roughly half of the tested benchmarks, especially in agentic tasks and coding, but trails in computer-use tasks and terminal commands.
Will Meta release more open models? Yes, the company announced plans to open the full weights of Muse Spark 1.2 in the coming weeks.
The article was prepared by the TechVisor team — practical IT media for people.
