AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

Muse Glimmer

Meta Superintelligence Labs
open weights / Overall score: 3.0

Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, distilled from Muse Spark and built to run agentic tasks locally on consumer hardware. A dense transformer with a 1.8-billion-parameter vision encoder, it reads text and images, writes text, holds a 131k-token context and takes a reasoning strength set per request. Meta also publishes 4-bit GGUF and ExecuTorch builds and a speculative-decoding drafter.

Scored apart from Muse Spark, the larger Meta model it is distilled from, which has its own entry.

Openness

3 high confidence
3.0
weights
open(BF16 safetensors on HF meta-models/Muse-Glimmer-30B, ungated, with first-party GGUF and ExecuTorch builds)
data
closed(the card names its sources only as public, third-party and Meta product data, and releases none)
code
partial(chat template and model configuration on the Hub only, no training or post-training code)
license
Apache-2.0(OSI, the LICENSE file in every Muse Glimmer repository)
usage-policy
conduct(USAGE_POLICY.md beside the weights lists prohibited uses and says the model is not intended for anyone under 18, and the Apache-2.0 LICENSE does not refer to it)

Muse Glimmer's weights download from Hugging Face without a gate under Apache-2.0, in full precision and in Meta's own 4-bit builds. Meta publishes neither the training data nor any training code, so the model can be run and fine-tuned but not rebuilt.

  • https://huggingface.co/api/models?author=meta-models&limit=50 recorded 2026-10-19

    The meta-models listing: Muse-Glimmer-30B, Muse-Glimmer-30B-GGUF, Muse-Glimmer-30B-ExecuTorch-PTE and Muse-Glimmer-30B-assistant all carry the tag license:apache-2.0, created 2026-08-09 and 2026-08-10. The two Muse-Spark-Open repositories in the same listing carry license:other.

  • https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B recorded 2026-10-19

    Hugging Face API record for meta-models/Muse-Glimmer-30B: `gated: false`, `private: false`, cardData license `apache-2.0`, files model-00001-of-00002.safetensors and model-00002-of-00002.safetensors beside LICENSE, USAGE_POLICY.md and chat_template.jinja. No dataset tag and no training scripts among the files.

  • https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/LICENSE recorded 2026-10-19

    LICENSE file: the Apache License text, "Version 2.0, January 2004", through "END OF TERMS AND CONDITIONS" and the standard appendix, with no added terms.

  • https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/README.md recorded 2026-10-19

    Raw model card: front matter "license: apache-2.0"; "Training Data: Multimodal content sourced from publicly available data, data provided by third parties and information from Meta's products and services"; the "Released Artifacts" table lists weights, two 4-bit builds, the DFlash drafter and the perception encoder, "All artifacts are released under Apache 2.0". No training code is linked.

  • https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/USAGE_POLICY.md recorded 2026-10-19

    "Muse Glimmer Usage Policy": "You should not use, or allow others to use, Muse Glimmer to" violate the law, assist military, weapons or critical-infrastructure uses, or deceive others, and "Muse Glimmer is not intended for individuals under the age of 18". No limit on commercial use.

Adoption

3 high confidence
3.0

Adoption is measured as Hugging Face downloads summed across Meta's full-precision, GGUF and ExecuTorch builds of Muse Glimmer, most of them for the GGUF build. The speculative-decoding drafter is left out of the sum, because it is an auxiliary checkpoint that cannot run the model on its own, not a way to get the model.

Capability

3 high confidence
3.0

Muse Glimmer scores well above most open-weight models of its size on the Artificial Analysis index, ahead of Mistral Medium 3.5, a model four times its size, and a little behind o3. That places it level with OpenAI's o-series.

  • https://artificialanalysis.ai/models/mistral-medium-3-5 recorded 2026-10-19

    "Mistral Medium 3.5 scores 14 on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 8)"; Intelligence rank "# 8 / 67".

  • https://artificialanalysis.ai/models/muse-glimmer recorded 2026-10-19

    "Muse Glimmer (High) scores 17 on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 8)"; Intelligence rank "# 14 / 141"; the embedded record gives intelligenceIndex 17.48, license "Apache 2.0" and weights at huggingface.co/meta-models/Muse-Glimmer-30B.

Verified 2026-10-19, except some scores, still awaiting re-confirmation