Meta introduces Muse Spark 1.1 with improvements to multimodal reasoning for agent and coding tasks

Meta has presented its updated artificial intelligence (AI) model Muse Spark 1.1, It comes with enhanced capabilities in multimodal reasoning and coding, designed to carry out agent tasks with advances in the use of tools and computers.

The team of Superintelligence Labs de Meta launched its new family of Muse AI models in April of this year debuting with Spark. Now, Meta has presented Muse Spark 1.1 as its evolution, with improvements that push “the limits of efficiency and performance” and that, together with its recently introduced imaging model, Muse Image, continues to propel the company to its vision of personal superintelligence.

Specifically, Meta has highlighted the enhanced capabilities of Muse Spark 1.1 in personal agent taskswhich require planning and managing a variety of external applications and services. Furthermore, you can tackle complex projects much more quickly than the previous version, being trained to “orchestrate multi-agent systems and optimize end-to-end latency.”

This was announced by the company in a statement on its blog, where it has specified that the new model, as the main agent, can gather context, make a plan, and delegate execution to parallel subagents.

Meta has also specified that Muse Spark 1.1 can manage your 1 million token context windowwith the ability to remember its previous actions, recovering information from other jobs and preserving the code of critical steps. As a result of all this, in third-party evaluations, stands out for surpassing Anthropic’s Opus 4.8 model a JobBench, MCP Atlas y DeepSearchQA.

Likewise, Muse Spark 1.1 also excels in IT workflows, adapting to changing requirements and with the ability to navigate through unknown interfaces with hardly any need for human intervention.

On the other hand, in the section of codingthe performance of the new model has also improved in real tasks involving “extensive and complex” code bases, as the company has assured. This means that it allows you to diagnose and correct errors, implement new functions in business systems and perform large-scale code migrations, among other tasks that Meta has exemplified.

In fact, Muse Spark 1.1 excels in use cases like creation of web applications, where it has significantly advanced compared to the original version in evaluations such as Vibe Code Bench v1.1 and SWE Atlas-Codebase QnA.

Regarding the multimodal reasoningMeta has ensured that it can interact with real environments and generate correct results, for example, it can describe images and videos in detail or generate “visual artifacts from code.”

In practice, the model can “analyze images and audio, preserve details throughout an extensive workflow, and use those details when operating computers on behalf of the user,” as the technology company explained.

Finally, the company has stressed that Muse Spark 1.1 is prepared for cybersecurity, with a “strong resistance” to direct jailbreaks and indirect attacks through untrustworthy data, prompt injection and attacks targeting developers.

All in all, Muse Spark 1.1 is now available via the Meta Model public preview API for developers in the United States, so they can test their capabilities.