AI Radar

Muse Glimmer

Available · Language, Multimodal, Reasoning

Muse Glimmer is a 30-billion-parameter open-weights model in Meta's Muse series, released on Hugging Face under the Apache 2.0 licence for self-hosting. It accepts text and images and reasons before answering; Meta says it was distilled from Muse Spark outputs and aims it at local agent workflows on a Mac or PC with a single consumer GPU. [1] [2] Primary source

Timeline of Muse Glimmer →

Claims and evidence

  • Released Primary source[1] [3]Model Release Date: August 2026
  • Status AvailableWeights downloadable from Meta's Hugging Face organisation meta-models (not gated) as of 2026-10-01. Primary source[3] [2]
  • Derived from (distillation) Muse Spark EditorialMeta says Muse Glimmer was trained on Muse Spark outputs with logit distillation but names the teacher only as Muse Spark, not a version; the link to the first Muse Spark record is an editorial choice. Primary source[1]How We Trained Muse Glimmer [2]Architecture [3]introduction
  • Change · Licensing First Muse model with published weights, under the Apache 2.0 licence; the Muse Spark models are offered only as hosted services.Compared with Muse Spark Primary source[1]
  • Change · Size A dense 30B model that Meta designed to run on a Mac or PC with a single consumer GPU.Compared with Muse Spark Primary source[1] [2]Model variants
  • Change · Architecture Ships with a DFlash-based speculative-decoding drafter and 4-bit quantized weights for local deployment.Compared with Muse Spark Primary source[3]Released Artifacts
  • Change · Context length Default context window of 128K tokens, with longer contexts supported.Compared with Muse Spark Primary source[2]Model variants
  • Input text, image Primary source[2]Model variants [3]Supported modalities
  • Output text Primary source[2]Model variants [3]Supported modalities
  • Feature Reasoning modeReasons before answering; reasoning strength low, medium, high or xhigh is set in the system prompt. Primary source[2]introduction [3]Best Practices: Reasoning Strength
  • Feature Function calling Primary source[1]Built for Agents
  • Feature MultilingualTrained on data from more than 100 languages. Primary source[2]Supported languages
  • Open weights YesApache 2.0 licence. Primary source[1]first paragraph [3]License; Released Artifacts
  • Context window 128K tokens by default; longer contexts supported tokensThe Hugging Face model card gives the context length as 131,072+. Primary source[2]Model variants [3]Context length
  • Parameters 30BThe Hugging Face model card gives about 29.6B including the vision encoder (a perception encoder of about 1.8B parameters). Dense architecture. Primary source[1]second paragraph [2]Model variants
  • Access Open-weights downloadSelf-hosted; not served on the Meta Model API. The blog said hosted access through partners such as Together AI, Fireworks AI and OpenRouter would follow in the coming days. Primary source[1]Get Started With Muse Glimmer Today [4]Muse Glimmer

Lineage

Predecessors

No known predecessor.

Successors

No known successor.

Based on

  • Muse Spark · 8 April 2026 · derived (distillation)

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

All model releases from Meta on AI Radar →