Meta releases Muse Glimmer, a 30B open model under Apache 2.0 built to run on one GPU
Originally announced by Meta . We link the primary source so you can read it for yourself.
Meta has released Muse Glimmer-30B, an open-weight multimodal model built to run on a single consumer graphics card. For a local AI audience, the most significant thing about it is not its size or its capabilities but its licence.
The news
Muse Glimmer-30B is a roughly 29.6-billion-parameter dense model paired with a vision encoder, so it accepts images as well as text, and it is aimed squarely at agentic use: tool calling, multi-step reasoning, and recovering from its own failures. It has a context window of 131,072 tokens, and Meta released the weights on Hugging Face under the Apache 2.0 licence, in several formats, including full-precision BF16 and 4-bit quantised variants, plus a separate drafter model for speculative decoding. The model is distilled from Muse Spark, a much larger closed model that Meta describes as sitting behind its consumer AI stack.
It is designed to run locally. Meta ships 4-bit quantised variants that it targets at 24GB and 32GB of VRAM, with claimed quality degradation of around 1.0% and 0.2% respectively against the full-precision model. Those figures are Meta’s own, and we treat them as claims pending independent testing, but the practical point stands: a 30B multimodal model that fits a single 24GB card is very much aimed at the hardware our readers actually own.
Why the licence matters
This is the part worth dwelling on. Meta’s Llama models, including its multimodal Llama 3.2 Vision and the widely used Llama 3.3 70B, ship under the Llama Community License, which permits commercial use but attaches conditions: a separate licence is required above 700 million monthly active users, and attribution and other terms apply. We have flagged those conditions repeatedly, because for some organisations they matter.
Muse Glimmer drops all of that. Apache 2.0 is a permissive, widely understood licence with no user thresholds and no attribution-style conditions on use, which puts this model on the same footing as the Qwen and Mistral open-weight families. For anyone who wanted a capable Meta model without reading the licence twice, that is a meaningful change.
The context
The release came with a lengthy essay from Mark Zuckerberg arguing for open-weight AI, framed partly as a way for models developed in the United States to compete with the strong open-weight models coming out of Chinese labs. Meta also said larger models are coming. Whether the new “Muse” name signals a broader move away from the Llama brand is not something Meta has spelled out, so we will not read more into it than that.
What is clear is the direction of travel, and it is not Meta alone. The shift toward genuinely permissive licensing has been one of the year’s throughlines, from Mistral moving its flagship family to Apache 2.0 to the Qwen and DeepSeek releases. A permissively licensed, locally runnable multimodal model from Meta is a notable addition to that pattern.
The technical angle
Two details are worth a note. Distillation is the practice of training a smaller model to mimic a much larger one, here the closed Muse Spark, which is how a 30B model can punch above the weight its size would suggest. And the separate drafter model enables speculative decoding, a technique where a small, fast model proposes tokens that the main model verifies in one pass, speeding up generation when the guesses are good. Both are aimed at making a capable model practical on modest hardware.
What to watch
Muse Glimmer is available now, with a confirmed licence and known memory requirements, so unlike some recent announcements this one is immediately usable. We have added it to the model catalogue; the benchmark and quantisation figures there are Meta’s own, pending independent testing. If the promised larger models arrive under the same licence, this becomes a more important release than a single 30B model would suggest.