Muse Glimmer 30B
Release in the Muse (Meta) family · version Glimmer-30B
Maintained by Meta (Meta Superintelligence Labs)16
Muse Glimmer is an open-weight model from Meta Superintelligence Labs with about 29.6 billion parameters, including a roughly 1.8-billion-parameter perception encoder. It is a dense transformer that takes text and images as input and produces text, and was distilled from Muse Spark for agentic tasks on consumer hardware. The model card lists a context length of 131,072+ tokens and a January 4, 2026 knowledge cutoff.1
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Model hub: Muse Glimmer collection (external site: huggingface.co)
- License: License (Apache 2.0) (external site: huggingface.co)
- Release notes: Meta developer blog post (external site: dev.meta.ai)
Availability and license
Overall availability
Weights can be downloaded from Hugging Face without a gated access request, under Apache 2.0. The repository also includes a separate Muse Glimmer Usage Policy.245
Availability is separate from permission: read the license before using or redistributing.
The repository's LICENSE file is the standard Apache License 2.0 text, and the model card says all released artifacts (BF16 weights, two 4-bit quantized variants, the DFlash drafter, and the perception encoder) are under Apache 2.0. The repository also includes a separate Usage Policy listing prohibited uses, such as military and weapons applications, which the model card links under responsible use. The LICENSE file does not reference that policy. The model card also says the model is not intended for people under 18.415
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Full-precision BF16 weights and two 4-bit quantized variants, plus the speculative-decoding drafter and the perception encoder, are published on Hugging Face without gating.123 |
| Inference codeIs code for running the model published? | Public | Runs in open-source runtimes rather than a Meta reference repository. Hugging Face's launch post documents use with Transformers, llama.cpp, and vLLM.71 |
| Training codeIs the code used to train the model published? | Unknown | Not assessed. |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The model card describes multimodal data from publicly available sources, third parties, and Meta's products and services, curated by vendors and Meta staff, covering more than 100 languages. The data is not released.1 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The model card says the model was distilled from Muse Spark and lists safety SFT and safety RL steps at a high level; no full recipe is published.1 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The model card reports benchmark and safety results, and Meta publishes an evaluation methodology document. Evaluation code was not located in this review.18 |
What it is useful for
The model card lists local AI agents, coding agents, tool use and function calling, multimodal reasoning over screenshots, charts, and documents, synthetic data generation, and LLM-as-a-judge evaluation, for commercial and research use.1
Run and use notes
- The model card states that quantizing the weights to about 4-bit precision brings the language model under 20 GB, and that together with the KV cache, perception encoder, and drafter it targets a 24 GB or 32 GB GPU memory envelope. The card lists 64 GB VRAM as the target hardware for full precision.1
- The model card reports speed measurements for the 4-bit K-Quant-17GB variant with the quantized DFlash drafter, run with llama.cpp on an NVIDIA RTX 5090 and with ExecuTorch on Apple M4 Max and M5 Max machines, at batch size 1.1
Organization context
Provenance and derivatives
The model card describes Muse Glimmer as distilled from Muse Spark, a Meta model whose weights are not public. Its perception encoder is a ViT-G/14 described in a paper that the model card links.1
- Derived from: Muse Spark (external site: ai.meta.com) — Teacher model for distillation; weights not public.
- Derived from: Perception Encoder (ViT-G/14) — Vision encoder cited in the model card (arXiv 2504.13181).
Other releases in the Muse (Meta) family
No other releases in this family have been assessed.
In the news
Meta releases Muse Glimmer as an open-weight model under Apache 2.0
In August 2026 Meta Superintelligence Labs published Muse Glimmer, a model of about 30 billion parameters that takes text and images as input and was distilled from Muse Spark. Its weights can be downloaded from Hugging Face without an access request under the Apache 2.0 license, and the repository also includes a separate Muse Glimmer Usage Policy that lists prohibited uses. Muse Glimmer has its own release record under the Muse family.
U.S. eligibility
Eligible · basis: U.S. headquarters
The model card names Meta Superintelligence Lab as author, and Meta's developer blog points to the meta-models Hugging Face collection as the download location, which confirms the repository is Meta's. Meta Platforms, Inc. has its principal executive offices in Menlo Park, California, per its Form 10-K.169
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.