Meta has released the weights for Muse Spark Open, a very large mixture-of-experts model aimed at coding agents, long-running tool workflows and multimodal reasoning. The official Meta model card dates the release to October 15, 2026. It lists 816 billion total parameters but about 42 billion active per token, with eight experts selected from a pool of 256. That distinction matters: sparse activation reduces work per token, not the size of the full checkpoint that an operator must store and serve.
The model accepts text, images, video and audio and produces text, with a documented inference context of up to one million tokens. Meta provides BF16 and NVFP4 variants and identifies deployment paths including SGLang, vLLM, Transformers, TensorRT-LLM, llama.cpp and MLX. The model card gives substantial memory requirements for self-hosting: at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16. It also describes a two-Mac-Studio M5 Ultra configuration. These are publisher deployment statements, not hardware tests by Open Model Weights.
The Hugging Face repository shows configuration files and a sharded Safetensors weight manifest, while access to the model remains gated. A listing of files is not proof that an unauthenticated user can download every shard. Teams evaluating the release should inspect the actual authorized artifact revision, supported modality processors, serving implementation and the operational cost of maintaining a long-context agent.
The license is a material part of this story. Meta's Muse Model License requires a visible 'Built with Muse Spark Open' notice. Its separate-agreement condition applies to qualifying model-as-a-service operators whose combined gross revenue exceeds US$20 million over a consecutive twelve-month period; the clause explicitly excludes certain internal and product-feature uses. The license also incorporates a usage policy. Meta publishes benchmark comparisons, but none is independently reproduced here. This is an important new open-weight option, not an unrestricted Apache-licensed release.
What this changes for the evidence layer
The existing Open Model Weights registry already contains separately field-verified BF16 and NVFP4 Muse Spark Open model profiles, each linked to an observed official repository revision and exact weight-file metadata. Both are explicitly marked gated: file listings do not establish unrestricted public download. The Muse Model License requires attribution/branding and can require a separate agreement for qualifying large model-as-a-service operators. Meta's performance figures remain publisher claims, not independent Open Model Weights measurements.
Meta Superintelligence Labs official Muse Spark Open model card and Muse Model License
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗
