Meta's Muse Spark Open is out: 816B open weights, $1.25/$4.25 on API hosts, and a two-Mac Studio local run

Meta released Muse Spark Open, an open-weights model with 816 billion total parameters, 42 billion active, and a 1 million token context. Meta's post calls it natively multimodal. AI at Meta lists text, image, video, and audio. Weights are on Hugging Face in BF16 and NVFP4. The BF16 model card says a local GPU needs at least 700GB of VRAM for NVFP4 and 2TB for BF16.

Meta says two Apple M5 Ultra Mac Studios, via MLX, hold the model. Mark Zuckerberg posted that it runs on the new Mac Studios with Apple's John Ternus. Meta's own demos, using Muse Code on the open model, ship iOS changes to Ice Cubes and drive DaVinci Resolve end to end. Those are Meta's demos.

The hosted rate is on the pricing table.

What it costs

OpenRouter lists $1.25 per million input tokens and $4.25 per million output. Its table shows that rate at Together, Nebius, Baseten, and Fireworks. Cache read is $0.15 per million on Together, Baseten, and Fireworks. Nebius does not print a cache price. Meta's post also names Groq and Modal as hosts.

Muse Wonder on Meta's own Model API is $4 input and $20 output. Claude Opus 5.5 is $4 and $20. Those are the published API stickers, not this model's.

The Muse Model License on the model card allows free commercial use if the product displays "Built with Muse Spark Open." A model-as-a-service seller whose gross revenue, with affiliates, exceeds $20 million in any twelve months has to make a separate deal with Meta before offering the model.

Should you switch

Try the hosted API if you want a Muse-class coding agent without Muse Code's $50 Power or $200 Maximum plan. Run it locally only if you have the card's GPUs or two M5 Ultra Mac Studios. Keep Claude or Grok for daily work until someone other than Meta publishes benches.