Meta Superintelligence Labs put Muse Spark Open on Hugging Face. It is 816 billion parameters total, 42 billion active per token, with a million token context window. This is the model. It is not NVIDIA DGX Spark hardware. It is not XHToken Spark-X2.5.
What shipped
Meta Superintelligence Labs released Muse Spark Open on October 15, 2026. MarkTechPost covered the announcement.1
The weights live on Hugging Face. The collection page lists the release.6 The BF16 checkpoint is at meta-models/Muse-Spark-Open-816B-A42B.3 The NVFP4 checkpoint is at meta-models/Muse-Spark-Open-816B-A42B-NVFP4.5
The model is an 816B MoE with 42B active per token. It uses 256 experts with 8 active per token. Context is 1M tokens. Input is text, image, video and audio. Output is text.
License is Muse Model License dated October 15, 2026. It permits use, fine-tuning, distillation and redistribution. Products must display "Built with Muse Spark Open". MaaS providers with more than $20M 12-month revenue need a separate agreement with Meta. Internal use, research and powering own product features are exempt from the revenue clause.
Formats are BF16 about 1.6 TB and NVFP4 about 530 GB. The model is also listed on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
What changed
The open weights drop is dated October 15, 2026. Meta's blog introduces it as "The Most Capable US Open-Weight Model to Date". That is a vendor claim.2
Native multimodal in with text out is new for this open release. The 1M token context sits on an 816B total, 42B active MoE. The NVFP4 cut about 530 GB sits beside the BF16 cut about 1.6 TB.
The license is Muse Model License with a MaaS revenue clause. It is not Apache or MIT.
MCP-Atlas 89.2 is top in Meta's table. JobBench 61.8. Terminal-Bench 2.1 76.4 trails Kimi K3 85.0, GLM 5.3 83.9 and Qwen3.8 82.0.
PyTorch blog coverage notes the release.4
Why it matters
NVFP4 checkpoint is about 530 GB. BF16 is about 1.6 TB.
It is a poor fit for single GPU home lab boxes. Even the NVFP4 cut is hundreds of gigabytes.
It is a good fit for labs that already run dual high memory Mac Studios or datacenter inference stacks, or for teams that will use a hosted API.
The license is Muse Model License, not Apache or MIT. Downstream engineers should read the MaaS revenue clause and the "Built with Muse Spark Open" display rule before local deploy.
Cloud listings on SageMaker, Model Garden and Foundry let some teams take the same weights without downloading 530 GB.
