Meta released the weights for Muse Spark Open, an open-weight coding model, and Amazon Web Services made it available in SageMaker JumpStart, the machine learning hub inside Amazon SageMaker AI.[1][2] The release lets teams run a model trained on real developer workflows such as multi-file refactors and long debugging sessions inside their existing cloud environment, Amazon Web Services said.[1]
Muse Spark Open uses a mixture-of-experts architecture with 816 billion total parameters, of which 42 billion are active for each token, accepts text, images, documents, audio and audio-video inputs, and returns text, according to the Amazon Web Services post.[1] It supports a context window of one million tokens, so a single prompt can include an entire project with dependencies and thousands of files, the post added.[1] SiliconANGLE reported that the two Mac Studios had about a terabyte of memory between them.[2]
SiliconANGLE reported that the model started from Muse Spark 1.2, which shipped in August, and that Meta then added safety post-training so the model holds up without the guardrails of Meta's own services.[2] TNW reported that Meta published Glimmer in August, a 30-billion-parameter model distilled from Muse Spark, under Apache 2.0.[3] Weights are also available on Hugging Face, Google Cloud Model Garden and Microsoft Foundry Models, and the model runs on inference engines including SGLang, vLLM and Apple MLX, Meta said.[4]
The weights carry the Muse Model License with limited commercial terms that allow free use for many developers but require a separate agreement once hosted services exceed $20 million in annual revenue, SiliconANGLE reported.[2] Military and critical-infrastructure uses are prohibited under the licence.[2]
