Meta Superintelligence Labs has released Muse Spark 1.1 — a multimodal reasoning model built for agent tasks. Its selling point isn't a single benchmark score, but packing "ultra-long context + multimodal perception + agent orchestration + coding" into a single model, with an OpenAI-compatible API to boot. The release also includes the Meta Model API public preview, which developers can plug into right now.

Core positioning

Muse Spark 1.1 is a model built for agent tasks. Its main capabilities include:

• Tool and computer use: it can operate desktop apps, call MCP servers, and use custom skills
• Coding: diagnosing complex bugs, implementing new features in enterprise-scale codebases, and executing large-scale code migrations
• Multimodal understanding: images, video, PDFs; vision-to-code and ultra-detailed image/video captioning

1 million-token context + automatic compression

This is one of Muse Spark 1.1's hardest specs: a 1M-token context window.

More importantly, the model can actively manage that 1M tokens:

• Remembering operations it performed earlier
• Retrieving information from work done long ago
• Automatically compressing context while keeping what later steps need

For agent workloads, this solves the common pain point of "the context blowing up halfway through a long task."

Agent orchestration: main agent + subagents

Muse Spark 1.1 is trained to serve as both a main agent and a subagent:

• As a main agent: gathering context, making plans, delegating execution to parallel subagents, and optimizing end-to-end latency
• As a subagent: focusing on its own task, understanding the available tools, and knowing when to escalate the problem back to the main agent

This multi-agent orchestration makes it more efficient than traditional computer-use models that just "click step by step" when handling complex workflows that span multiple apps with dynamically changing information. The model decides when to write a script to automate something, when to click through the UI directly, and when to generate operations in batches.

Computer use

Muse Spark 1.1 is specifically optimized for computer-use scenarios that span multiple apps with information changing in real time:

• Maintaining context across long sessions
• Adapting to changing requirements
• Handling unfamiliar interfaces with minimal human intervention

Meta gave a concrete example: a Facebook Marketplace listing agent — you shoot a short video with your phone, and the model automatically extracts useful photos, analyzes the product, operates your browser, and publishes the listing on Marketplace for you.

Coding

On real-world, complex codebase tasks, Muse Spark 1.1 improves substantially over the first generation:

• Diagnosing and fixing complex bugs
• Implementing new features in enterprise systems
• Executing large-scale code migrations

It supports mainstream agentic coding toolchains, including common features like planning mode, goal conditioning, subagent delegation, and context compression. In one Meta demo, the model built a chat web app inside OpenCode: it automatically took screenshots, spotted a user-visible fault, traced it back to the relevant code, then fixed it and verified the fix.

💡 Tip: Replit CEO Amjad Masad's verdict on Muse Spark — "One-million-token context, full multimodal support (images/video/PDF), built-in search and citations, strong reasoning, top-tier coding (especially frontend and design), structured outputs, and parallel tool calls, all packed into a clean OpenAI-compatible package. A complete agent foundation."

How to use it

• Meta AI app / meta.ai: available right now in "Thinking" mode
• Meta Model API (public preview): developers can officially integrate through an OpenAI-compatible interface
• Third-party coding tools: Replit, Cline, OpenClaw, and others are already integrated

Safety

Meta ran a complete evaluation under its Advanced AI Scaling Framework covering three categories: chemical/biological, cybersecurity, and loss-of-control risk. The results show Muse Spark 1.1 sits within the safe range across all frontier risk categories, with strong resistance to jailbreak attacks, indirect prompt injections, and developer prompt attacks, plus lower hallucination and sycophancy rates.

Who it's for

• Agent application developers: 1M context + multi-agent orchestration suits long-chain automation
• Coding tool teams: OpenAI-compatible API + subagent delegation means low integration cost
• Agents that need multimodal perception: scenarios combining visual/audio understanding with real action (like operating a browser on the user's behalf)

References

• Introducing Muse Spark 1.1 (Meta AI official blog)
• Meta AI Research