Muse Glimmer
Meta's 30B dense multimodal model for local agents, distilled from Muse Spark, Apache 2.0, and quantized to fit a 24 GB GPU.
In short. Muse Glimmer is Meta's smaller Muse model, built to run on your own computer instead of in the cloud. It reads text and images, can use tools, and is meant for coding, document work and personal assistants. A compressed 17 GB version fits on a 24 GB graphics card, and the license is Apache 2.0, so you can use it commercially.
Summary
Meta released Muse Glimmer on 10 August 2026 under the Apache 2.0 license. Meta Superintelligence Lab describes it as a 30-billion-parameter model distilled from Muse Spark and built for autonomous agent tasks on consumer hardware. The card gives about 29.6 billion parameters in total, including a roughly 1.8-billion-parameter perception encoder that reads images.
It is a dense transformer, not a mixture of experts: 52 layers, with a repeating pattern of three local sliding-window attention layers followed by one global layer. It takes text and images, writes text, has a context length of 131,072 tokens or more, and was trained on data from more than 100 languages. Reasoning effort is adjustable.
Meta designed it to run locally. The BF16 weights are about 60 GB. Meta also ships compressed builds: a "K-Quant-Dynamic" build aimed at 32 GB of memory with 0.2% average degradation, and a 17 GB build aimed at 24 GB with 1.0% average degradation, measured across 15 benchmarks. A small DFlash draft model speeds up generation; Meta reports 233 tokens per second with it on an RTX 5090, against 75 without. These figures are Meta's own.
Meta's benchmark table compares it with Gemma 4 31B and Qwen3.6-27B and shows it ahead on several agent benchmarks such as MCP Atlas and DeepSearch QA, but behind on others. Treat the table as the vendor's view and test on your own tasks. Day-0 support was announced for Transformers, llama.cpp and vLLM.
Good for
- Private, on-device agents for coding, document analysis and personal assistants
- Running on a single 24 GB or 32 GB GPU, or a well-equipped Mac, with the quantized builds
- Reading screenshots, charts and documents alongside text
- Commercial products, since Apache 2.0 has no revenue or hosting conditions
Watch out for
- Images are supported, but audio and video are not
- The 128K context is much shorter than the million-token windows of its larger sibling
- The quantization-loss figures and speed numbers are Meta's own measurements
- It is distilled from a stronger model, so on hard tasks it will trail models many times its size
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Glimmer 30B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Glimmer 30B BF16 | 29.6B | 68 GB | 34 GB | 20 GB | Gaming GPU |
| Glimmer 30B GGUF (17 GB K-quant) | 29.6B | 68 GB | 34 GB | 24 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 24 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source models for coding and AI agents
- Best open-source multimodal models in 2026 (images, video and audio)
- Meta releases Muse Spark Open, an 816B open-weights model
Frequently asked questions
- What is Muse Glimmer?
- Muse Glimmer is Meta's smaller Muse model, built to run on your own computer instead of in the cloud. It reads text and images, can use tools, and is meant for coding, document work and personal assistants. A compressed 17 GB version fits on a 24 GB graphics card, and the license is Apache 2.0, so you can use it commercially.
- Can I use Muse Glimmer commercially?
- Yes. Muse Glimmer is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Muse Glimmer open source?
- Muse Glimmer is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Muse Glimmer on my own computer?
- Yes, on a good graphics card. The smallest variant, Glimmer 30B, needs about 20 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Muse Glimmer released?
- Muse Glimmer was released by Meta on August 10, 2026.
- How long a context window does Muse Glimmer have?
- Muse Glimmer supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Muse Glimmer model card (Hugging Face) huggingface.co
- Hugging Face blog: Meta is back with Muse Glimmer (10 Aug 2026) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Muse Spark Open Meta | 816B 42B active | 1M | Muse Model License (Conditional) | 563 GB GPU server | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Mage-VL-Cowork Microsoft | 4.74B | 32K | Apache 2.0 (Permissive) | 3.27 GB Laptop or phone |