AI Trend Research
EN

← October 15, 2026 briefing · added in the 04:00 KST update

P0 Model Open weights Language modelsMultimodalAgents04:00 updateMetaHugging FaceNVIDIAAWSApple

Meta releases 'Muse Spark Open', an 816B MoE open-weight model with 42B active and 1M-token context for coding and agents

Announced October 15, 2026

Verified (fact)

Meta has released an open-weight model called 'Muse Spark Open'. It uses an MoE (Mixture of Experts) architecture with 816B total parameters and 42B active per token, takes text, image, audio and video as input, and generates text. Context reaches up to 1,048,576 tokens across all modalities combined. It was trained for agentic coding tasks such as multi-file editing, debugging and tool use, and checkpoints are provided in BF16 and NVFP4 formats. Hugging Face Transformers v5.20.0 supported the model on release day, and it was also listed on Amazon SageMaker JumpStart. NVIDIA published instructions for running it on GB300 NVL72, and an Unsloth version appeared in the community. According to Meta's announcement, it is also releasing Helion GPU kernels for SGLang, and the model underwent additional safety post-training before being evaluated under its 'Superintelligence Scaling Framework'.

Needs fact-check (claim)

Meta claims it is "the most capable American open-weight model released to date." According to The Verge, Meta said it is "the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework," and that it worked with Apple to improve performance on Macs.

Why it matters (interpretation)

After launching Muse Wonder as a closed frontier model, Meta is again releasing a frontier-class open-weight model. This could shift the competition between U.S. open-weight models and large Chinese MoE models (StepFun, Ling and others). With day-one support from AWS, NVIDIA, Hugging Face and Apple MLX, enterprises can readily deploy it on their own infrastructure. The 'most capable' claim, however, needs to be confirmed by independent benchmarks.

Sources

More about Meta