Meta Unveils Muse Spark 1.1 AI Model With 1 Million-Token Context Window, Model API Preview

Meta Platforms on Thursday announced Muse Spark 1.1 as its latest multimodal reasoning model designed for agentic AI tasks. The company claims it brings significant improvements over its predecessor when it comes to tool use, coding, computer interaction, and multimodal reasoning. Developed by Meta Superintelligence Labs, the AI model supports a context window of up to one million tokens. Alongside, Meta also released a public preview of the new Meta Model API for developers.

Muse Spark 1.1 Available to Developers via Meta Model API

The company says that Muse Spark 1.1 is available to developers via the new Meta Model API that is now in public preview. It is claimed to be designed to perform complex agentic tasks that require planning, reasoning, and coordination across multiple applications and external services. The AI model can generalise to new native tools, Model Context Protocol (MCP) servers, and custom skills without any additional training, Meta said in a blog post.

The company highlighted the improvements in task orchestration, with Muse Spark 1.1 being capable of serving as a primary agent that gathers context, formulates plans, and delegates work to multiple subagents running in parallel. Compared to the original Muse Spark model, its architecture is claimed to deliver faster completion of projects while reducing end-to-end latency.

Meta is also touting Muse Spark's one million-token context window, which helps it manage long-term memory, improve information retrieval from previous conversations, and its ability to compress the context while retaining key details for future use.

Other areas of improvement include computer-use capabilities and coding. As per the company, Muse Spark 1.1 can navigate desktop apps, interact with unfamiliar software interfaces, and take a call on when tasks should be automated using scripts. It also promises better performance during enterprise-scale software engineering tasks like debugging large codebases, implementing new features, executing large-scale code migrations, and building complete web apps. Meta claimed that its developers are already internally using the model for software development and AI research workflows.

Benchmark results shared by Meta show its competitiveness against other AI models across several industry evaluations. The model is claimed to have 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 front, Meta claims to have conducted safety evaluations under its Advanced AI Scaling Framework before deployment. It is said to have remained within acceptable threshold margins across chemical, biological, cybersecurity, and loss-of-control risk categories. Further, the AI model is claimed to have demonstrated improved resistance to jailbreak attempts, prompt injection attacks, and hallucination compared to the original model.