Meta Muse Glimmer open weights return with a 30B local agent model
Meta has released Muse Glimmer’s 30B weights under Apache 2.0, reopening its local-model strategy while keeping Muse Spark proprietary.
By Dominic Okoye · Staff Writer
· 3 min read
Meta released Muse Glimmer on August 10, making the 30-billion-parameter model’s weights available under the Apache 2.0 license. The Meta Muse Glimmer open weights release gives developers a new local model for agent workflows after the company’s proprietary Muse Spark launch had raised doubts about its commitment to the developer community built around Llama.
Meta says Glimmer is designed to run continuously on a Mac or PC with a single consumer GPU, targeting local agents, coding, function calling and model-as-a-judge tasks. The company says the model can process interleaved text and images through a dedicated perception encoder. It trained Glimmer by distilling outputs from Muse Spark, its larger proprietary model.
What is Meta Muse Glimmer and what hardware does it need?
Glimmer is an open-weight language model, meaning Meta has released the downloadable model weights for others to use, deploy and modify under its stated license terms. That differs from Muse Spark, which DeepLearning.AI reported was released as a closed model, leaving its weights under Meta’s control.
The practical constraint is memory. Meta says the full-precision version requires more than 55 GB of memory. Its roughly 4-bit quantized language model is below 20 GB, and Meta says a 24 GB or 32 GB memory setup can accommodate the model alongside its working memory, image encoder and draft model. The Register separately reported that a 4-bit version reduces the weights to just under 16 GB, though usable local deployment still requires capacity for more than the weights alone.
Meta also uses speculative decoding, in which a smaller draft model proposes blocks of tokens for the main model to verify. Meta claims this improves generation speed without reducing output quality. Those are company performance claims, and operators evaluating the model for production work should test it against representative tasks rather than rely on benchmark comparisons alone. A practical model-evaluation process should include the relevant inputs, tool calls and failure cases.
A return to weights, not a release of Meta’s top model
The release is a partial reversal, not confirmation that Meta will publish the weights for its most capable systems. Meta’s own benchmarks compare Glimmer with Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B. The Register characterized it as a smaller local-model option for small and midsize businesses and enthusiasts, rather than a challenger to the largest open-weight systems from Chinese developers.
Alexandr Wang, Meta’s chief AI officer, said an open-weight version of Muse Spark 1.2 would arrive “soon.” He provided no release date, parameter count or licensing terms in the statement reproduced by Digg.
The launch also arrives amid a policy argument over the U.S. position in open-weight AI. Reuters reported that Mark Zuckerberg called for lower policy barriers for domestic open-weight developers instead of restrictions on foreign models. He said U.S. labs face additional limits on training data. For enterprise teams, local weights create an option for customization and on-device deployment, but Glimmer’s release alone does not settle Meta’s longer-term approach to its frontier models.
This story draws on original reporting from The Register.
