Meta launches Muse Spark Open, an 816B-parameter coding model trained on NVIDIA GB300 NVL72 systems

Meta/Muse models

Meta launched Muse Spark Open on October 15, 2026, a coding-optimised mixture-of-experts model. NVIDIA's technical blog says Meta trained it on GB300 NVL72 Blackwell Ultra systems for agentic coding, and it ships as NVFP4 weights on Hugging Face and as a downloadable NVIDIA NIM.

• 816B total parameters with 42B active per token, top-8 routing across 256 experts, and a 1M-token context window.
• Agent features include web search, subagent coordination, and computer use across desktop, mobile and browser.
• NVIDIA's vendor figures: over 50K tokens per second per GPU and over 600 tokens per user, in a test on four Blackwell GPUs with speculative decoding.
• Supported on the SGLang, vLLM and TensorRT-LLM inference stacks; fine-tuning runs through NVIDIA NeMo AutoModel and Megatron Bridge.