Meta Superintelligence Labs has released Muse Spark Open, an 816-billion-parameter Mixture-of-Experts (MoE) open-weight model with a 1-million-token context window. The native multimodal model processes text, image, video, and audio inputs to generate text, activating 42 billion parameters per token across its 62 layers.
Meta claims Muse Spark Open is the most capable U.S. open-weight model, delivering top benchmark scores in tool use and knowledge reliability, though it lags behind leading Chinese open models like Kimi K3 and GLM 5.3 in coding tasks. Weights are distributed in BF16 and NVFP4 formats on Hugging Face, with quantization enabling local deployment on setups as small as two Apple M5 Ultra Mac Studios.
The release includes day-one support for runtimes such as vLLM, SGLang, and MLX, along with hosted API deployments across Baseten, Groq, and OpenRouter. The permissive Muse Model License allows commercial use, fine-tuning, and distillation without revenue restrictions for internal applications.
Why it matters
Provides startups and enterprises a massive open-weight multimodal foundation model for complex long-horizon agentic workflows.
Quantized NVFP4 builds enable local workstation deployment, reducing dependence on expensive proprietary cloud APIs.
Maintains competitive open-weight options against leading international models under permissive commercial licensing.
Source: marktechpost.com



