On October 15, 2026, Meta released the weights for Muse Spark Open, a natively multimodal MoE model with 816 billion parameters, of which 42 billion are active per token, and a 1 million token context window. For the first time for a model of this class, local deployment without the cloud is claimed: two Mac Studios with M5 Ultra are combined via MLX. The weights are uploaded to Hugging Face in BF16 and NVFP4 formats and are served by seven API platforms; Meta calls Muse Spark Open the most capable open model in the US, but there are no independent assessments of this claim yet.

image
image
image

What happened

The release is the publication of weights in open access: two variants, BF16 and NVFP4, are uploaded to Hugging Face, and hosting is also provided on Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry. Inference is open from day one on seven API platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Supported runtimes include SGLang, vLLM, TensorRT-LLM, llama.cpp, MLX, and Unsloth, with Meta bringing custom Helion kernels for NVIDIA and AMD in SGLang. The model itself takes text, images, video, and audio as input and generates text. Along with the weights, the company demonstrated the Muse Code agent: it finds and fixes errors in the code of the iOS app Ice Cubes and manages editing in DaVinci Resolve, viewing video frame by frame.

Context

Muse Spark Open is built on a mixture of experts scheme, where only a portion of the parameters is activated for each token — here about 5% — which allows combining a large total knowledge capacity with a moderate cost per request. Native multimodality means that all types of input data are processed by a single network, not an ensemble of separate models, and this simplifies agentic pipelines where documents, screenshots, and media files are processed simultaneously. The wide context window enables analysis of long codebases, multi-hour recordings, and frame-by-frame work with video, on which agent interaction scenarios with real applications are built. The release fits into the general shift of recent years: labs are increasingly moving the frontier level of capabilities from closed APIs to open weights. The distinction of this release is a ready-to-use package for immediate community verification: two weight formats, six runtimes, and availability on third-party platforms on the day of release.

Why this matters for the industry

For the industry, the main thing is that the frontier agentic level is coming to open weights for the first time with a ready path to production: teams can connect the model via partner APIs without their own hardware or deploy it themselves on vLLM and SGLang, gaining control over request cost and data privacy. Optimizations have already entered the main stack: Meta's custom Helion kernels in SGLang speed up inference by 1.03–3.95 times on NVIDIA/AMD and provide a 3% end-to-end improvement on AMD MI350. By opening such a level of capabilities on the day of release, Meta removes two barriers at once — inference cost and privacy requirements, since sensitive data can be processed locally or on-premise. The value for startups shifts up the stack, to vertical products, proprietary data, and distribution, while the margin of pure API resellers will be under pressure. If independent assessments confirm the claimed capabilities, price pressure on closed APIs should be expected in the coming months.

Why this matters for users

The model can be tried today and without your own hardware: a free or cheap entry is open via OpenRouter, Baseten, and other hosters from the partner list. The second path is self-hosting: download BF16 or NVFP4 from Hugging Face and run vLLM or SGLang on your own GPU server. The third is fully local operation: two Mac Studios with M5 Ultra are combined via MLX, weights are pulled directly from Hugging Face, and with up to 512 GB of memory per machine and a bandwidth of 1.2 TB/s, the entire model lives on the desktop without the cloud, which is suitable for sensitive data. On this basis, MVPs are already being built: an agent on top of a desktop application, automatic code patching in the style of Muse Code, or analysis of long audio and video recordings. Before commercial deployment, it is worth checking the terms of the Muse Model License: the license contains restrictions that are better to know in advance.

What is still unknown / limitations

Meta's claim of leadership among open models in the US in the release materials is not supported by an evaluation methodology or comparison numbers with competitors, so until independent community checks appear, it should be perceived as a marketing formulation. The Muse Code demonstrations are vendor-selected showcase examples without a description of tasks and metrics, which does not allow separating real capabilities from well-chosen scenarios. There is no data on latency and inference cost on real traffic in the sources, so it is too early to judge production readiness; the correct path now is pilots on your own traffic with your own tests. The sharding scheme of the local scenario is not disclosed: in BF16, the 816 billion parameter model takes up about 1.6 TB and does not fit into the total 1 TB of memory of two Mac Studios, so it is likely that the NVFP4 variant or more aggressive compression is used, and the exact configuration needs to be checked in practice. Finally, the commercial restrictions of the Muse Model License require separate legal review before embedding the model into a product.

Sources

Author

Look at AI, editorial team