Meta unveils 816B-parameter Muse Spark Open
Meta's open-weight Muse Spark Open activates 42 billion parameters per token and supports one million tokens of context, with text, image, video and audio input.
Meta unveiled Muse Spark Open, an open-weight AI model with 816 billion parameters, the company said on October 15, 2026, according to VentureBeat. The model uses a mixture-of-experts architecture that activates only 42 billion parameters per token.
Muse Spark Open supports up to one million tokens of context and accepts text, image, video and audio as input, producing text-only output. Meta presented a setup running the model locally on two M5 Ultra Mac Studio machines with 512GB of unified memory each.
According to Meta, Muse Spark Open outperformed competing models on the JumboBench and MCP-Atlas tool-use evaluations, and scored 39 on the Artificial Analysis Intelligence Index. On terminal coding and long-horizon software engineering tasks, however, it trailed Moonshot AI's Kimi K3 and GLM 5.3.
The license permits execution, modification, fine-tuning and redistribution. Companies with more than $20 million in annual revenue, including affiliates, must sign a separate agreement with Meta to use the model in a service business. Together AI set API pricing at $1.25 per million input tokens and $4.25 per million output tokens.
Separately, the Financial Times reported that Nvidia is weighing an additional investment in or an outright acquisition of open-source AI startup Reflection AI. The talks are at an early stage, and an acqu-hire arrangement — hiring staff and buying only technology usage rights — is also under consideration.
Hardware requirements
Running Muse Spark Open locally requires substantial memory. Apple plans to release a 512GB memory product at the end of October and has not disclosed its price. A Mac Studio with the required M5 Ultra chip and 256GB of memory currently costs at least $10,799. In March, Hugging Face said in an evaluation that Chinese models accounted for 41% of downloads over a year, surpassing US models.
Quick answers
How many parameters does Muse Spark Open have?
Meta's Muse Spark Open has 816 billion parameters and uses a mixture-of-experts architecture that activates only 42 billion parameters per token.
What does Muse Spark Open cost via API?
Together AI set API pricing at $1.25 per million input tokens and $4.25 per million output tokens.
Can companies use Muse Spark Open commercially?
The license permits execution, modification, fine-tuning and redistribution, but companies with more than $20 million in annual revenue, including affiliates, must sign a separate agreement with Meta to use the model in a service business.