Meta plans to fund its Muse AI agent through subscriptions rather than advertising as it rolls out in Japan, Nikkei Asia reported. The agent passed 10 million US downloads in its first 40 days, AI Affairs reported. The plan emerged in an interview with Meta AI head Alexandr Wang, published on 24 October Japan time. Japan is the starting point for Meta’s international expansion of Muse, which the company says will remain free of advertising.
Key points
- Meta plans to use subscriptions rather than advertising for its Muse agent as it expands into Japan.
- Wang is positioning the service against Chinese AI models while Meta renews its work on open-weight models.
- Meta’s separate Muse Spark Open model has a licence provision requiring certain service providers to make a separate agreement above $20 million in revenue over 12 months.
Muse takes a subscription route into Japan
The Japanese launch gives Meta an international market for the Muse agent and a way to charge users directly for it. Wang linked the subscriber approach to competition from Chinese AI models in the Nikkei Asia interview. Meta says the agent will stay ad-free.
The choice places a different source of revenue behind Muse from the advertising business associated with Meta’s existing products. An earlier AI Affairs report on Meta’s 26% revenue rise also covered Mark Zuckerberg’s Muse growth and cost target. The Japanese plan concerns the agent sold to users, rather than the terms under which developers can use Meta’s downloadable model.
Wang joined Meta last year. Under his leadership, the company has shifted its attention back towards open-weight development, which allows others to download and adapt a model’s parameters. Muse Spark Open, released on 15 October, is part of that effort and is distinct from a subscription to the Muse agent.
Muse Spark Open faces Chinese models
Meta released Muse Spark Open with 816 billion parameters, according to SQ Magazine. The publication reported that 42 billion parameters are active for each request and that the model accepts a context of one million tokens. Meta built it for coding and longer-running agent tasks.
Chinese developers have been prominent in open-weight models, with Alibaba’s Qwen, Moonshot AI’s Kimi K3 and Z.ai’s GLM among the contenders, Agence France-Presse reported. Meta claims Muse Spark Open rivals some Chinese alternatives and outperforms them on certain tasks. Those claims concern model performance, rather than the number of people willing to pay for the Muse agent.
SQ Magazine reported a score of 39 for Muse Spark Open on the Artificial Analysis Intelligence Index, against 46 for Xiaomi’s MiMo-V2.6-Pro, using a figure Meta cited for its own model. The publication said Artificial Analysis had not listed Muse Spark Open at the time, leaving Meta’s cited score unconfirmed. Its comparison gives the Chinese competition a measure beyond Meta’s own description of the model.
Apple hardware and the Muse model licence
Meta has also worked with Apple to run Muse Spark Open on two Mac Studio computers. The arrangement uses Apple’s M5 Ultra chips and MLX software framework, PPC Land reported. Meta demonstrated coding work on the machines, while developers can obtain the model through cloud platforms as well. AI Affairs previously covered the model’s release for Mac Studios.
Running the downloadable model locally keeps users’ prompts away from Meta, PPC Land reported. That arrangement applies to the model running on a user’s own computers, rather than establishing a condition of the Muse subscription service. The two products therefore offer different ways to use Meta’s AI work: one through an agent account, the other through model weights that developers can deploy.
The Muse Model License permits commercial products and modified versions of Muse Spark Open, SQ Magazine reported. Its separate commercial threshold applies to AI service providers earning more than $20 million over 12 months from the model: they must notify Meta and sign another agreement. The publication reported that internal use, research and embedded products are exempt from that provision.
