A Major Leap for Meta's Agentic AI
Meta Superintelligence Labs has officially launched Muse Spark 1.1, a substantial upgrade to its multimodal reasoning model. Designed from the ground up for agentic workflows, the new release brings meaningful improvements in tool and computer use, coding, and multimodal understanding, pushing the performance-efficiency frontier further than its predecessor.
The launch arrives alongside Meta Image and marks another step toward the company's stated vision of "personal superintelligence" — AI that can plan, create, and take action on a user's behalf across everyday digital tasks.
Alongside the model itself, Meta is opening a public preview of its new Meta Model API, giving developers direct access to Muse Spark 1.1 for the first time. The model is also live in "Thinking" mode inside the Meta AI app and on meta.ai.
Built to Orchestrate, Not Just Respond
One of the standout upgrades in Muse Spark 1.1 is its ability to handle complex, multi-step agentic projects far faster than before. Rather than working through tasks linearly, the model is trained to orchestrate multi-agent systems, minimizing end-to-end latency by delegating work intelligently.
- As the main agent, it gathers context, forms a plan, and delegates execution to parallel subagents.
- As a subagent, it stays focused on its assigned task, understands the tools available to it, and knows when to escalate a decision back up to the main agent.
- It also zero-shot generalizes to new native tools, MCP servers, and custom skills — no retraining required.
Supporting all of this is a 1 million token context window that the model actively manages. Muse Spark 1.1 can recall actions from much earlier in a session, retrieve relevant information on demand, and compact its working memory in a way that preserves the critical steps needed for later work.
Smarter Computer Use
Muse Spark 1.1 shows notable gains in computer-use tasks that span multiple applications, especially when information changes mid-task. It's built to maintain context across long sessions, adapt to shifting requirements, and navigate unfamiliar software interfaces with minimal hand-holding.
Instead of clicking through every step one at a time, the model has learned when to automate versus when to interact directly with an interface — writing scripts for speed when appropriate, clicking through UI when that's simpler, and batching multiple actions together at each step. In one demonstration, the model organized a dinner party order and automatically adjusted the plan when new information came up mid-task, without needing user intervention.
Coding Gains on Real-World Codebases
Coding performance saw a substantial boost, particularly on tasks involving large, complex, enterprise-grade codebases. Muse Spark 1.1 can diagnose and fix intricate bugs, implement new features, and execute large-scale code migrations.
The model was trained to work smoothly across diverse agentic coding harnesses, reliably handling multi-turn dynamics like planning mode, goal conditioning, subagent delegation, and context compaction. In a demo using OpenCode, Muse Spark 1.1 built a chat web app, took automated screenshots to catch user-visible bugs, traced them back to the responsible code, implemented fixes, and validated the results — blending coding, multimodal understanding, and tool use in a single workflow.
Internally, Meta says the model significantly outperforms the original Muse Spark on its primary coding benchmark, Meta Internal Coding Bench, and is now competitive with other leading models in the space. Researchers are even using Muse Spark 1.1 to automate parts of their own model development and evaluation pipelines.
Stronger Multimodal Reasoning
Beyond code, Muse Spark 1.1 shows real strength in perception and multimodal reasoning — particularly in visual-to-code generation, detailed image and video captioning, and executing agentic workflows that require understanding both visuals and audio.
This matters most when perception and action need to happen together. In one example, the model reasoned over smartphone video footage to extract useful product photos, then used a browser autonomously to create a Facebook Marketplace listing on the user's behalf.
Safety First
Before deployment, Meta says it ran Muse Spark 1.1 through extensive safety evaluations under its Advanced AI Scaling Framework, which sets thresholds across frontier risk categories including chemical and biological risks, cybersecurity, and loss of control. According to Meta, the model operates within safe margins across all of these categories, and shows improved resistance to jailbreaks, prompt injection, and indirect attacks from untrusted data — alongside lower hallucination rates and reduced sycophancy.
Industry Reaction
Early partners testing Muse Spark 1.1 through the new Meta Model API have praised its combination of long-context handling, strong reasoning, and coding ability. Replit CEO Amjad Masad described it as a complete agentic foundation that packs massive context, full multimodal support, and strong coding performance into a single OpenAI-compatible package. Cline's Saoud Rizwan highlighted its strong tool use at a price point that makes real coding workloads viable at scale, while Box's Yashodha Bhavnani noted its competitiveness with frontier models on enterprise workflows in sectors like professional services and public sector operations.
What's Next
With Muse Spark 1.1 now available in public preview via the Meta Model API, developers can start building agentic applications on top of it today. Meta says this release is just one step in an ongoing research push, with more capable models already in training.
