Meta Launches Muse Spark 1.1 AI Model with Agent Task Capabilities

Meta’s Superintelligence Lab has released the Muse Spark 1.1 model, which excels in Agent capabilities, enabling it to act as a primary Agent that assigns tasks to multiple sub-Agents for parallel processing. It supports tool invocation, computer operations, and code generation. With a context window of up to 1 million tokens, the model’s API pricing is set at $1.25 per million tokens for input and $4.25 per million tokens for output—offering competitive pricing. The model can also understand video content and automatically complete tasks such as listing second-hand goods. Meta is also advancing its in-house AI chip, Iris, scheduled for production in September, with the goal of reducing compute costs. General users can access the model directly through the Meta AI app.

Article author and source: 36Kr

Meta AI has become a bit unfamiliar.

Today's international front-page news in the AI world was even more intense than usual.

Following the well-received Grok-4.5 and with GPT-5.6 looming large, Meta’s Superintelligence Lab (MSI) has just unveiled its new model: Muse Spark 1.1.

Fun fact: This promotional “essay” comes three years after his last post on X, and his previous post before that was back in 2012.

Releasing a new model from Meta at this critical juncture means it must deliver real substance.

Meta didn’t come empty-handed—on the contrary, they rolled out two major updates right away: Muse Spark 1.1 has undergone a significant upgrade, and the Meta Model API has also launched into public beta, allowing developers to directly access this flagship model for the first time.

The benevolent figure who built half of the industry’s reputation on open-source models has finally openly launched an API for sale. Meta has framed this upgrade as “pushing the frontiers of performance and efficiency,” stating that the simultaneous release of the model and API brings them one step closer to their vision of “personal superintelligence.”

Help you pursue your goals, create what you’ve imagined, strengthen your relationships, and take action on what matters most to you.

First, let’s look at the key focus of this upgrade: Agent capabilities.

According to Meta, Muse Spark 1.1 is a multimodal reasoning model designed for agent tasks, with enhancements across key areas including tool calling, computer operations, code generation, and multimodal understanding.

It is specifically trained to handle "individual agent tasks": such as planning, scheduling, and executing actions across multiple external applications and services. These tasks are often challenging due to their lengthy workflows, numerous variables, and dispersed sources of information—the model must not only answer questions but also manage and coordinate tasks effectively.

The most interesting aspect of Muse Spark 1.1 is that it was trained to function as a "contractor."

After receiving a complex task, it acts as the primary agent: gathering context, formulating a plan, assigning tasks, and delegating different subtasks to multiple parallel sub-agents to reduce end-to-end processing time.

Conversely, when it becomes a sub-agent itself, it knows how to cooperate: complete its assigned tasks, recognize available tools, and promptly report any limitations rather than working blindly.

The context window also supports 1 million tokens and employs an active management mechanism. It doesn’t just accommodate more content—it remembers what it has done, retrieves information from much earlier, and retains key details needed for subsequent steps when compressing the context.

Muse Spark 1.1 can also be used immediately in a zero-shot manner for new tools, MCP services, and custom skills that users have never encountered before.

Regarding computer use, Meta's approach is very practical.

In the past, many computer operations required agents to repeatedly view the screen, re-interpret the situation, and re-click, making the entire process painfully slow.

Muse Spark 1.1 has learned to choose the appropriate method based on the scenario: write a script directly if it’s faster, use the interface if it’s more convenient, and when batch execution is needed, it can generate and execute multiple steps at once.

Meta provided an example of "organizing a dinner gathering": during the meal ordering process, when the user temporarily changes the conditions, the model independently identifies the new circumstances and adjusts its plan, requiring no repeated input from the user.

Muse Spark 1.1 can maintain context across applications, long workflows, and continuously changing information—scenarios that previously caused agents to lose track. It also rarely requires manual guidance when encountering unfamiliar interfaces.

In terms of coding capabilities, Meta this time emphasized real-world, large-scale codebase scenarios.

For example, diagnosing and fixing complex bugs, adding new features to enterprise-grade systems, and executing large-scale code migrations. Muse Spark 1.1 also shows significant improvements over the first generation in use cases such as creating web applications and end-to-end question answering.

It has also been trained to be highly adaptable.

It can adapt to common practices across different coding toolchains, harnesses, and agent coding suites, including planning modes, goal condition setting, sub-agent delegation, and context compression.

In other words, it can quickly get up and running in any coding tool.

Meta demonstrated an OpenCode debugging demo: the model first built a chat web application, then automatically took a screenshot, identified issues visible to the user from the screenshot, traced the relevant code based on those findings, made corrections, and continued validating the changes.

Writing code, reviewing screenshots, adjusting tools, and verifying results—this entire workflow closely mirrors the actual working method of an Agent in real-world development.

Within Meta, engineers and researchers are already using Muse Spark 1.1 extensively. According to official statements, in the internal code evaluation benchmark Meta Internal Coding Bench, version 1.1 shows significant improvement over the initial version and is now capable of competing directly with leading models.

Even more ironically, researchers have already begun using it to automate model development and evaluation workflows.

In another demo, Muse Spark 1.1 evaluated itself on a subset of DeepSWE tasks using varying levels of reasoning intensity, and then generated an analytical dashboard based on the results.

Take the exam yourself, grade it yourself, and issue your own report card (doge).

In terms of multimodality, Meta is primarily promoting "watch and work simultaneously."

In other words, Muse Spark 1.1 excels not only in understanding images, videos, and audio, but also in performing real-world tasks after comprehending this content.

It can interact with real-world environments and produce fact-based outcomes. Vision-to-code, fine-grained image and video description, and multimodal agent workflow execution are capabilities explicitly highlighted by Meta.

In real-world scenarios, it can view images and videos, listen to audio, retain these details over long processes, and then use this information to operate a computer.

The most down-to-earth demo comes from Facebook Marketplace.

A user records a short video of a product with their phone; the model extracts usable photos from the video, identifies the product category and condition, then automatically opens a browser to list the used item on behalf of the user.

Notably, Muse Spark 1.1 is also Meta’s second Muse family product released this week. On Wednesday, Meta unveiled Muse Image, an image generation model previously codenamed Mango, primarily designed for creators and advertisers.

While promoting image models to attract creators and advertisers, Meta is also advancing its Agent programming model to draw in developers and enterprise customers. Meta’s AI product line is now showing clear commercial specialization.

On the API front, several of Meta's early partners have already come forward in support.

The overall evaluation of Muse Spark 1.1 can be summed up in one sentence: It is a comprehensive agent foundation, combining long-context processing, strong coding, powerful reasoning, and tool-use capabilities sufficient to support large-scale agent workloads.

Replit CEO Amjad Masad believes the most impressive aspect is that Meta has bundled so many capabilities into a single model: million-token context, support for images, videos, and PDFs, built-in search with citations, strong reasoning, structured outputs, parallel tool calling, and top-tier coding ability.

In his view, Muse Spark 1.1 excels particularly at frontend and design tasks. He also mentioned in passing that the entire suite of capabilities is packaged in a "clean, OpenAI-compatible format."

Cline CEO Saoud Rizwan emphasized price.

In his view, Meta is clearly targeting "serious agent coding": it has strong tool-calling capabilities and has driven prices down to a level suitable for running coding tasks at scale. The combination of high capability and low cost is uncommon, which is why Cline wants developers to start using it as soon as possible.

In fact, the input price for Muse Spark 1.1 is $1.25 per million tokens, and the output price is $4.25 per million tokens. Alexandr Wang, head of Meta AI, told CNBC in an interview that this pricing is “very aggressive and highly attractive.”

Meta’s ability to drive prices to this level is certainly not just due to its model alone. According to an internal memo reviewed by Reuters, Meta plans to begin producing its own AI chip, Iris, starting in September this year.

Iris is part of Meta's MTIA project, short for Meta Training and Inference Accelerators. As planned, it is one generation in Meta's fourth-generation AI chips, designed to enhance the AI systems underlying core products such as Facebook and Instagram.

The memo noted that Iris spent approximately six weeks in the bug testing phase without discovering any major issues. Broadcom is the design partner for this chip, and TSMC is responsible for manufacturing.

Developing its own chips is crucial for Meta: it reduces computing costs and decreases dependence on third-party chip suppliers like NVIDIA and AMD.

However, Iris is not positioned as a replacement for GPUs.

According to the memo, it will complement Meta’s continued large-scale purchases of Nvidia and AMD GPUs. In other words, Meta is now pursuing a multi-pronged approach: continuing to buy top-tier GPUs while accelerating the development of its own chips.

And the chip is just the first layer. The same memo also mentions that Meta is advancing a larger-scale infrastructure expansion: 7 GW of computing capacity scheduled to come online in 2026, growing to 14 GW by 2027.

This scale is no longer merely an expansion of a conventional data center—it is an energy and computing infrastructure project for the era of superintelligence.

Oh, by the way, although Meta has started selling APIs, that doesn’t mean it’s completely abandoning open source. Wang said in the interview that Meta remains committed to open source, and an internal variant of Muse Spark is currently under development with plans to open source it in the future. However, he did not disclose any specific release timeline.

Regular users also don’t need to wait. According to Meta’s official announcement, Muse Spark 1.1 is now available in the Meta AI App and meta.ai, and can be used directly in 'Thinking' mode.

Regarding security, Meta states that it has completed extensive pre-deployment evaluations in accordance with its Advanced AI Scaling Framework, which is primarily designed to define evaluation methodologies, threat models, and deployment thresholds for state-of-the-art models.

Complete security evaluation details were documented by Meta in the "Muse Spark 1.1 Evaluation Report." Detailed performance evaluation data are also included in the report.

At the end of the blog, Meta stated that even more powerful models are already in training, with multiple releases occurring within a single week and additional models still in reserve. At least in terms of its approach, Meta’s Superintelligence Lab no longer intends to remain low-key.