Muse Glimmer: Meta's weights under Apache 2.0
On 10 August 2026 Meta published the weights of Muse Glimmer 30B under the Apache 2.0 licence, together with quantised versions from the maker itself. The card gives about 29.6 billion parameters including the vision encoder, a context length from 131,072 tokens, and a purpose: autonomous agentic tasks on consumer hardware without reaching for the cloud.
Why it matters
Open weights for a capable multimodal model under a permissive licence arrived for the second time that month, and for the first time not from China. Apache 2.0 allows commercial use without a separate negotiation, and the size together with the maker's own quantised builds makes running it on your own hardware an ordinary act rather than an exercise.
The model card, signed by Meta Superintelligence Lab, gives about 29.6 billion parameters, including a frozen ViT-G/14 perception encoder of roughly 1.8 billion; a tokenizer of 200,000 BPE tokens plus 2,048 special ones; up to 4,096 visual tokens per image; a knowledge cutoff of 4 January 2026; and distillation from Muse Spark. The model ships with a DFlash drafter that proposes blocks of 16 tokens in a single pass for the main model to verify. The day comes from the repository's commit history: the repository was created on 9 August, the card written that evening, the weights added in a commit at 07:51 UTC on 10 August and the licence and usage policy at 08:15 UTC. The card itself names only the month. What the record does not claim. Any of the comparisons in the card: the tables against Gemma4-31B and Qwen3.6-27B are the maker's own measurements, and no independent check was read as of the day of this record. That the model leads its class: in several rows of those same tables it does not. That the weights come without conditions: besides the licence, the repository carries a separate usage policy.