Muse Spark Open: 816B parameters and 700 GB VRAM to self-host
Meta released Muse Spark Open, an open-weight 816B-parameter MoE model with 42B active parameters and a 1M-token context. The NVFP4 build needs at least 700 GB of VRAM and BF16 needs 2 TB, while hosted API access costs $1.25 per million input tokens.
- 816B total parameters, 42B active, 256 experts, 1M-token context
- BF16 weights take 1.64 TB, NVFP4 takes 533 GB; card lists 700 GB VRAM
- OpenRouter API: $1.25 per 1M input and $4.25 per 1M output tokens
- Selling it as a service past $20M revenue requires a Meta agreement
Read next
AI