Meta opens Muse Spark weights for local deployment
Meta has released Muse Spark Open, a model that accepts text, images, audio and video while producing text. Its weights give developers a route to run and adapt the model themselves, subject to the Muse Model License. The 816 billion parameter system activates only part of its network for each token, but its complete weights still need substantial memory.
Artificial Intelligence··Evening
Meta releases a model developers can run themselves
Meta has released Muse Spark Open, its model for handling text, images, audio and video while generating text. The weights are distributed through Hugging Face, a platform for sharing machine-learning models. The release lets developers download the model for their own infrastructure. It follows a collaboration with Apple focused on local operation, while use of the weights is governed by the Muse Model License.[1], [2]
Meta describes the intended work as extended coding and tool use: reading repositories, changing multiple files and debugging. Its demonstrations include software changes through Muse Code, the company’s coding service, and selecting and cutting video inside DaVinci Resolve, an editing application. These examples come from the company’s demonstrations of the newly released model.[1], [2]
Active experts still require the complete weights in memory
The model contains about 816 billion parameters, with approximately 42 billion active for each token, a unit of input or output. Its mixture-of-experts architecture selects eight experts out of 256 in each expert block. The other experts remain in memory. Hugging Face puts the higher-precision BF16 checkpoint at about 1.5 TB, distinguishing the stored model from the smaller portion used in each computation.[1]
Its context configuration allows 1,048,576 positions, shared by text and other inputs. The audio encoder turns incoming sound into tokens, while the perception encoder handles visual material. The resulting inputs occupy the same context space. For a full context in BF16, the technical post estimates about 266 GB of key-value cache, the memory that retains previously processed context.[1]
Serving and adaptation tools accompany the release
Software support includes Transformers, llama.cpp, SGLang, vLLM and MLX, tools for loading and serving models with differing hardware and input support. A fine-tuning example freezes the experts and trains smaller attention-layer adapters. Hugging Face describes that example as a run across 64 H100 graphics processors for 500 steps. Separately, Meta says its work with Apple allows operation on two M5 Ultra Mac Studio computers using MLX.[1], [2]
