[Digital Today intern reporter Seungah Yoo] Meta has unveiled Muse Spark 1.3, the most powerful large language model (LLM) it has released so far. The company said it has narrowed the performance gap with major artificial intelligence (AI) companies such as OpenAI and Anthropic.

On Sept. 3 (local time), Meta began providing Muse Spark 1.3 through the Meta Model API for paid use by developers, according to foreign media outlets including SiliconANGLE. It is currently available through Meta's coding tool Muse Code and the Meta Model API, and the company plans to apply it later to social media platforms such as Facebook and Instagram as well as to Meta AI.

Alexandr Wang (알렉산드르 왕), Meta's chief AI officer (CAIO), said in an interview with Bloomberg that Muse Spark 1.3 marks the biggest performance improvement among the models Meta has introduced. He said it is competitive with Anthropic's Claude Fable 5.1 by emphasizing coding and agent-style automation features, and claimed it outperforms OpenAI's GPT-5.6 Sol in code generation. He added that it is also ahead of current AI models from China.

Still, AI model comparisons can vary depending on evaluation methods and task types, making it difficult to judge superiority based only on companies' own claims. According to independent AI evaluation group Artificial Analysis, Muse Spark 1.3 Max, which is being made available on a limited basis to partners, scored 62 on the Intelligence Index. That is a higher score after Claude Fable 5.1 and Claude Opus 5. Another reasoning version of Muse Spark 1.3 currently available, called 1.3 xhigh, scored 61, the group said, putting it at the same level as GPT-5.6 Sol.

Meta has released four Muse Spark models over the past 5 months, accelerating the AI race. It has invested large sums in AI infrastructure and development to catch up with Anthropic, OpenAI and Chinese rivals. Last year, Meta CEO Mark Zuckerberg (마크 저커버그) spent more than $14 billion (about 19.0 trillion won) to acquire a stake in Scale AI, Wang's former company, and recruited Wang to lead the AI strategy at Superintelligence Labs.

Meta's large AI investments are also raising concerns among investors. With the timing of any return on investment uncertain, Meta has shifted away from its open-source-focused strategy carried over from its existing Llama models and is pivoting toward a proprietary model strategy closer to Anthropic and OpenAI.

Meta first introduced paid use for developers when it released Muse Spark 1.1 in July. At the time, Zuckerberg said he planned to make the Muse Spark family one of the cheapest LLMs on the market. The usage fees for Muse Spark 1.3 are the same as those for version 1.2 released in August.

Wang said Muse Spark 1.3 uses about 25 percent fewer tokens than the previous version to perform the same tasks. Meta said it also reduced tool calls by about 20 percent. It can handle multiple workflows simultaneously in a single long conversation, and its ability to follow complex and lengthy instructions and maintain context across multiple tasks has improved.

Safety has also been strengthened. The model has been improved so it can better understand the tasks it can perform and its limitations, and it is designed to ask for user confirmation before taking irreversible actions.

Meta has not yet provided specific details on whether it will release the weights for Muse Spark 1.3, but it has said it will release an open-weights version in the future. It is also developing a next-generation model called Watermelon that is larger and more powerful than Muse Spark. Zuckerberg has teased the release of Watermelon, but has not disclosed a specific timeline.

We’re excited to release Muse Spark 1.3 with improved performance on agentic and coding tasks, and a focus on real-world usability. Key capabilities: → Sustains longer-horizon work across multiple workflows in a single thread → More actively collaborates with users: it asks… pic.twitter.com/1GJUaFqrTZ