Meta Platforms released the next-generation large language model Muse Spark 1.1 to strengthen its position in the AI agent and code generation market

According to Zhitong Finance APP, just two days after launching the AI image generation model Muse Image, Meta Platforms (META.US) has once again upgraded its artificial intelligence (AI) product matrix. On Thursday, Meta released the next-generation large language model Muse Spark 1.1 and opened the API for public preview for the first time, further intensifying its competition in the AI agent and code generation market against rivals like OpenAI and Anthropic. Following the announcement, Meta's decline narrowed to 0.5%, having previously dropped by about 4%.

Alexandr Wang, head of Meta's Super Intelligence Lab, stated that Muse Spark 1.1 is Meta's "most powerful model in AI agents and programming tasks" to date. Compared to the initial Muse Spark released in April, which only provided API previews to a select few partners, the new version will be accessible to more developers through the Meta developer platform, allowing users to apply for a spot on the waiting list to gradually gain API access. However, Meta currently restricts API access to its own ecosystem and has not yet opened it to third-party model platforms like OpenRouter.

It is understood that Muse Spark 1.1 focuses on enhancing code generation and AI agent capabilities. Wang mentioned that the reason for prioritizing programming capabilities is that code generation is the core foundation for AI agents to complete complex, multi-step tasks, helping to create AI assistants that can autonomously execute work.

In terms of pricing, Meta has adopted a more competitive strategy. Wang stated that each new API account will receive a $20 free credit, with subsequent input tokens charged at $1.25 per million tokens and output tokens at $4.25 per million tokens, making the overall pricing more attractive compared to similar products from OpenAI and Anthropic, aimed at encouraging large-scale usage by developers.

Earlier this week, Meta had just released the AI image generation model Muse Image and integrated it into core products like Instagram, WhatsApp, and Meta AI, while also planning to apply it to advertising creative generation tools. The continuous launch of new Muse series products reflects Meta's acceleration in promoting AI commercialization in response to Wall Street's concerns about the returns on its massive AI infrastructure investments.

In recent years, Meta has been continuously expanding its AI capital expenditures, but compared to OpenAI, Anthropic, and Google (GOOGL.US), the company still lags behind in popular AI models and applications. Meanwhile, Meta has yet to establish a mature cloud computing business, although the company has indicated plans to enter this market in the future.

It is noteworthy that Meta's AI strategy is also undergoing changes. Previously, the company primarily relied on the open-source Llama series models to expand its developer ecosystem; now it is gradually shifting towards providing commercialized API services for self-developed closed-source models to explore new revenue sources However, Wang emphasized that Meta has not abandoned its open-source strategy. The team is still developing an open-source version of Muse Spark, but the release date has not yet been announced.

In addition, he revealed that Meta is currently training a next-generation AI model codenamed "Watermelon," which will further enhance performance, but the specific release date has not been disclosed. He also mentioned that he has begun using Muse Spark in scenarios such as health management, assisting decision-making by searching online information, reading academic papers, and integrating personal health data, believing that this is one of the important application directions for AI agents in the future