01 What happened
On October 15, 2026, Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model with approximately 816 billion parameters and about 42 billion active per token, accepting text, images, audio, and video as input while generating only text output.
02 Key details
- Muse Spark Open is an open-weight MoE model from Meta with 816B total parameters and 42B active per token; each MoE layer uses 256 experts and the router selects 8 per token.
- The model accepts text, images, audio, and video inputs, has a 1M-token context window, and generates text output only. Meta describes it as the most capable US open-weights model.
- The BF16 checkpoint is about 1.5 TB and runs on 4 nodes of 8 H200 GPUs; the NVFP4 checkpoint runs on a single B200 node. Fine-tuning with TRL and LoRA on attention layers is possible while experts stay frozen.
- Reasoning cannot be turned off in Muse Spark Open; the processor supports effort levels low/medium/high/xhigh. The model supports tool calling for text-only and multimodal inputs.
03 Why it matters
This release gives developers access to a large-scale open-weight multimodal model optimized for agentic coding, with concrete deployment and fine-tuning recipes across multiple inference engines.
04 Who it matters to
AI developers, AI agent developers, artificial intelligence researchers, machine learning engineers and voice interface developers.
Original source