Meta opens Muse Spark 1.1 to AI coding tools through a new API

In July 2026, Meta introduced Muse Spark 1.1, a new version of its in-house AI model, along with the Meta Model API for connecting it to development tools. The company says the model handles complex programming, bug fixing and agentic workflows better.

What the model gets

Muse Spark 1.1 is the second generation of Meta's Muse Spark, the company's first in-house model since it returned to the AI race. The first version came out in April 2026; Meta took developer feedback into account when preparing the update. The company calls the changes a qualitative leap over the first generation, but the available description gives no benchmarks, model size, request pricing or independent comparisons with competitors.

• Meta introduced the first Muse Spark model in April 2026.
• The update is called Muse Spark 1.1 and is available through the Meta Model API.
• The stated feature is finding and fixing complex bugs in code.
• The model is built for agentic scenarios, including systems with several agents.
• It can natively process images, video and documents.

Version 1.1 matters for more than its list of claimed features. Meta is moving the model from a standalone announcement into an integration layer: developers can connect it to AI coding tools through the API. That format lets teams embed the model in their existing environment instead of forcing them to work in a separate Meta interface.

What changes for developers?

According to Meta, Muse Spark 1.1 is meant to handle more complex programming tasks, including detecting and resolving complex bugs. This is not simply about generating a function snippet from a text prompt: the model is intended for cases where it must understand a problem, propose a fix and work that fix into the development process.

Meta places particular emphasis on end-to-end agentic scenarios across different applications. In such scenarios, an AI system receives a task, breaks it into steps, interacts with tools and passes results between specialized agents. Meta explicitly mentions multi-agent systems. That means Muse Spark 1.1 is positioned as a component of larger automation, not just a chat for writing code.

Native multimodal perception broadens the inputs. According to the company, the model can work not only with text and code but also with images, video and documents. For engineering teams, that is potentially useful for analyzing an error screenshot, a specification in a document, or visual material tied to a task. However, Meta has not described which formats are supported, what limits apply, or how accuracy was measured for each data type.

Why the API bet matters

The Meta Model API makes Muse Spark 1.1 available to AI coding tools. This shifts the point of competition: it is not enough for a company to release a strong model; it must become a convenient supplier for code editors, agent platforms and internal business tools. The easier a model is to connect and evaluate in a real project, the better its chance of entering the everyday development stack.

Meta already has its own model lineup and large infrastructure, but in this announcement the company stresses Muse Spark's practical role. The wording about complex bugs and multi-agent systems targets a market where models are judged by work completed, not just by the quality of answers in conversation. The stated capabilities remain Meta's own position: the publication contains no independent verification or comparable numerical results.

For developers, the practical next step is to test the model on their own repositories, typical bugs and review processes. It is especially important to assess the quality of the fixes, the ability to retain context between steps, and the cost of running an agentic chain. Without this data, it is impossible to judge how competitive Muse Spark 1.1 really is at programming tasks.

What this means

Meta is turning Muse Spark from an early launch into a product for integration with AI development tools. If the stated capabilities hold up on real projects, the new model could compete not only in code generation but also in automating the full cycle of engineering tasks.

*Meta has been designated an extremist organization and is banned in Russia.