Meta Platforms announced on Thursday the release of Muse Spark 1.1, its newest multimodal reasoning model aimed at agentic AI applications. The firm asserts notable enhancements over the prior version in areas such as tool utilization, programming, computer interaction, and multimodal reasoning. Created by Meta Superintelligence Labs, this AI model features a context window reaching one million tokens. In addition, Meta launched a public preview of its new Meta Model API for developers.

Muse Spark 1.1 Available to Developers via Meta Model API

The company states that Muse Spark 1.1 reaches developers through the new Meta Model API now in public preview. It is designed to handle complex agentic tasks involving planning, reasoning, and coordination across multiple applications and external services. The AI model can generalize to new native tools, Model Context Protocol servers, and custom skills without further training, according to a company blog post.

Meta emphasized gains in task orchestration, noting that Muse Spark 1.1 can act as a primary agent to gather context, create plans, and assign work to multiple subagents operating in parallel. Its architecture is said to enable quicker project completion and lower end-to-end latency compared with the original Muse Spark model.

Meta also highlights the one million token context window, which supports long-term memory, better information retrieval from past conversations, and context compression while preserving essential details for later use.

Additional improvements cover computer-use abilities and coding. According to the company, Muse Spark 1.1 can navigate desktop applications, engage with unfamiliar software interfaces, and decide when to automate tasks via scripts. It also delivers stronger results on enterprise-scale software engineering work such as debugging large codebases, adding new features, performing large-scale code migrations, and constructing full web apps. Meta noted that its internal teams already rely on the model for software development and AI research workflows.

Benchmark results shared by Meta indicate competitive performance against other AI models across multiple industry tests. The model reportedly achieved an MCP Atlas score of 88.1, JobBench score of 54.7, Humanity's Last Exam score of 62.1, and Finance Agent v2 score of 57.2. On coding benchmarks, it recorded 80.0 on Terminal-Bench 2.1 and 53.3 on DeepSWE 1.1.

On the security side, Meta says it performed safety evaluations under its Advanced AI Scaling Framework before deployment. The model stayed within acceptable threshold margins across chemical, biological, cybersecurity, and loss-of-control risk categories. It also showed stronger resistance to jailbreak attempts, prompt injection attacks, and hallucination compared with the original model.