Meta releases Muse Spark Open, an 816B open-weight agentic coding model

Meta calls Muse Spark Open the most capable US open-weight model to date and positions it for long-horizon tasks, complex software engineering and multi-step reasoning. The headline specs are 816B total parameters with 42B active per token, a 1M-token context window, and native input across text, image, video and audio. NVIDIA's companion post adds the architecture detail: top-8 routing across 256 experts, a 1.8B perception encoder for the non-text modalities, and NVFP4 checkpoints that Meta generated for Blackwell Ultra hardware, with a NIM container for deployment. Hugging Face is hosting the weights and has published guidance for running and fine-tuning the model across the main inference stacks. Two cautions apply. Meta's capability claim is its own, and independent index scores were not yet in today's evidence; Artificial Analysis rated NVIDIA's new Nemotron 3.5 Super VL the leading US open-weight model before Muse Spark Open was measured. The throughput figures in NVIDIA's post, up to 50K tokens per second per GPU, are vendor numbers on GB300 NVL72 systems most teams do not have. For teams that want a self-hosted backbone for coding agents, the practical question is whether a 42B-active model with a 1M window fits their serving budget better than a hosted frontier API.