Meta released Muse Spark 1.1 on July 9, 2026, opening the multimodal reasoning model to outside developers through a new public preview of the Meta Model API, and pairing the launch with a stress-test safety report covering chemical and biological risk, cybersecurity, and loss of control.

The model, built by Meta Superintelligence Labs, is described in Meta's own announcement as a multimodal reasoning system built for agentic tasks, with the company citing major gains in tool use, computer use, coding, and multimodal understanding compared with the earlier Muse Spark. It arrives days after Meta's separate launch of Muse Image, and the two releases together form what Meta frames as a step toward what it calls personal superintelligence, though that phrase describes a long-term company vision rather than any claim being made about Muse Spark 1.1 itself.

Joel Kaplan, Meta's Chief Global Affairs Officer, announced the developer opening on LinkedIn, writing that the company was "opening Muse Spark 1.1 to developers through the Meta Model API - high intelligence at one of the best price points in the market." Kaplan added that Meta had stress-tested the model before release "to make sure it's secure" and was "publishing the full results alongside the model, covering what we tested, what we found, and how we mitigated the risks." He framed the release within a broader argument about national competitiveness, stating that "US leadership in AI means giving people access to capable, secure models."

What changed in the model

According to Meta's announcement, Muse Spark 1.1 can actively manage a context window of 1 million tokens, a capacity the company says lets the model remember earlier actions, retrieve information from much earlier in a session, and compact its working memory in a way that preserves the steps needed for later work. A context window of that size determines how much text, code, or instruction history a model can hold in a single working session before it must summarize or discard earlier material - a technical constraint that has shaped how developers structure long agentic workflows across the industry.

The model's agentic capabilities extend to orchestration. Muse Spark 1.1, according to Meta, zero-shot generalizes to new native tools, Model Context Protocol servers, and custom skills, meaning it can work with previously unseen external tools without additional training specific to those tools. As a main agent, the model can gather context, form a plan, and delegate execution across parallel subagents; as a subagent itself, Meta states it adheres to an assigned task, understands the tools available to it, and knows when to escalate a problem back to the coordinating agent. The company says this multi-agent orchestration is designed to optimize end-to-end latency on complex projects, allowing the system to tackle work Meta describes as significantly faster than its predecessor could manage.

Computer use is a second area Meta highlighted. The model, according to the announcement, excels at workflows unfolding across multiple applications where information changes during the session, maintaining context across extended sessions and adapting to evolving requirements while navigating unfamiliar interfaces with what Meta describes as minimal human intervention. Rather than reasoning through each desktop interaction step by step, Meta says it trained Muse Spark 1.1 to judge when writing an automation script is faster than direct clicking, and to generate batches of actions within a single step when that approach is more efficient. Meta's own demonstration material describes an "agentic dinner party organization" example in which the model adjusts a food order automatically when new context arises mid-task, without requiring the user to intervene.

Coding gains and internal benchmarks

Coding performance improved substantially on real-world tasks involving large, complex codebases, according to Meta's announcement, with the model able to diagnose and fix complex bugs, implement new features inside enterprise-grade systems, and execute large code migrations. The company says Muse Spark 1.1 shows large gains over its first model in use cases such as building web applications and end-to-end question answering.

Meta states that the model was trained to adapt smoothly to diverse coding harnesses and to reliably handle complex multi-turn dynamics, and that it performs well with popular agentic coding setups supporting planning mode, goal conditioning, subagent delegation, and context compaction. A debugging demonstration described in the announcement shows Muse Spark 1.1 building a chat web application inside OpenCode, taking automated screenshots to identify failures visible to users, tracing those issues back to the relevant code, implementing fixes, and validating the changes - combining coding, multimodal understanding, and tool calling within a single workflow.

On Meta's primary internal coding evaluation, which the company calls Meta Internal Coding Bench, Muse Spark 1.1 significantly improves upon Muse Spark and is described by Meta as competitive with leading alternatives, though the announcement does not publish a comparative numerical score against specific named competitor models. Meta also states that developers and researchers across the company are using Muse Spark 1.1 daily, and that researchers have begun automating model development and evaluation tasks by incorporating the model into their own workflows, including a demonstrated example where Muse Spark 1.1 evaluates itself on a subset of DeepSWE tasks and produces an analysis dashboard from the results.

Multimodal perception and tool use

Muse Spark 1.1 also carries strengths in perception, multimodal reasoning, and tool use, according to Meta, with particular capability in visual-to-code artifact generation, what the company describes as ultra-descriptive image and video captioning, and agentic workflow execution across multimodal use cases. The company states that these multimodal capabilities matter most when perception and action need to happen together - inspecting visual and audio input, preserving details across a long workflow, and then using those details while operating a computer on a user's behalf.

One demonstration Meta cites involves a Facebook Marketplace agent: using video shot from a smartphone, the model extracts useful photographs, reasons about the product shown, and then operates a user's browser to create a Facebook Marketplace listing on that person's behalf.

Safety testing and the Advanced AI Scaling Framework

Meta says it conducted extensive safety evaluations before deployment, following what the company calls its Advanced AI Scaling Framework, which the announcement describes as defining evaluations, threat models, and deployment thresholds specifically for Meta's most advanced models. Across three frontier risk categories the company names - chemical and biological, cybersecurity, and loss of control - Meta states its evaluations show Muse Spark 1.1 operates within safe margins.

The model, according to Meta, demonstrates strong resistance to direct jailbreak attempts and to indirect attacks originating from untrusted data, prompt injection, and developer-prompt attacks. Meta reports that this translates into better adversarial robustness, lower hallucination rates, and reduced sycophancy relative to its earlier model. The company says its full safety posture for the 1.1 release is documented in a separate Muse Spark 1.1 Evaluation Report published alongside the model.

Availability and industry reaction

For the first time, Meta states, developers can begin building with Muse Spark 1.1 through the new Meta Model API, which entered public preview alongside the model's release. The model is also available now in what Meta calls Thinking mode inside the Meta AI app and at meta.ai.

Several early partners offered assessments included in Meta's own announcement. Amjad Masad, chief executive officer of Replit, is quoted saying: "What's most impressive about Muse Spark is how much it packs into one model: massive million-token context, full multimodal support (images, video, PDFs), built-in search with citations, strong reasoning, top-tier coding abilities (particularly frontend and design), structured output, and parallel tool calling - all in a clean OpenAI-compatible package. A complete agentic foundation." Replit's platform is used by enterprise teams including Zillow, Atlassian, Adobe, Accenture, and Databricks to build internal tools, and the company has previously integrated Anthropic's Claude computer use capability into its own agent product, as PPC Land reported.

Saoud Rizwan, chief executive officer of Cline, said: "Meta is clearly building for serious agentic coding - strong tool use at a price point that makes it viable to run real coding workloads at scale. That combination is rare, and it's exactly why we wanted Cline developers to have access early."

Yashodha Bhavnani, vice president of AI products at Box, said that "when tested against Box's enterprise work evaluation set, Muse Spark delivered enterprise capabilities competitive with today's leading frontier models," adding that the combination of that intelligence level with strengths in "structured, procedural workflows across industries such as professional services, public sector, and industrial operations" made the model, in her assessment, "a compelling choice for organizations."

Dave Morin, of the OpenClaw Foundation, offered a shorter assessment: "Muse Spark 1.1 is an awesome model for running agents. Fast, powerful, and fun with OpenClaw."

Meta's announcement closes by describing the release as "a testament to our research momentum," stating the company has additional, more capable models currently in training.

Context for the marketing and advertising industry

Muse Spark 1.1 is not an advertising product, and Meta's announcement makes no reference to Ads Manager, campaign automation, or measurement infrastructure. Its relevance to marketing and advertising professionals is structural rather than direct, and it sits inside a pattern that PPC Land has tracked closely across Meta's developer platform over the preceding months.

On April 29, 2026, Meta launched Ads AI Connectors and a companion Ads CLI, opening its advertising system to Claude and ChatGPT with write access to campaign creation, ad sets, ads, and creatives. Roughly two months later, Meta published a considerably more conservative Developer Tools MCP server, keeping nearly its entire surface area read-only, with webhook subscription management as the sole write capability across ten available tools. Read together with Muse Spark 1.1's zero-shot generalization to Model Context Protocol servers, the sequence describes a company building a foundation model with native orchestration ability at the same time it is deciding, tool by tool, how much of its advertising infrastructure that kind of model - whether Meta's own or a competitor's - should be allowed to touch.

That caution has a regulatory backdrop. Meta consolidated five previously separate developer compliance processes into a single system, a move PPC Land examined in detail in the context of the Developer Tools MCP launch. A model capable of orchestrating subagents across a company's tools, as Meta describes Muse Spark 1.1 doing, raises the practical stakes of getting that compliance layer right, since the party initiating an action may increasingly be an autonomous system rather than a person clicking through an interface.

Meta's own advertising business supplies the financial motive for this direction. The company's advertising revenue reached 58.1 billion US dollars in the fourth quarter of 2025 alone, according to PPC Land's coverage of Meta's fourth-quarter results, with the company stating it expects to consolidate more ranking models in 2026 than in the previous two years combined. Engineering productivity increased 30% during 2025 according to that same reporting, with the majority of the gain attributed to agentic coding adoption that accelerated toward the end of the year - the exact category of internal use Meta's Muse Spark 1.1 announcement now describes with specific internal benchmark claims.

The release also lands inside a broader industry-wide race to ship agentic, long-context coding models, one where rival announcements have arrived on a near-monthly cadence throughout the first half of 2026. Google's Gemini 3.5 series launched at Google I/O on May 19, 2026, with the company's own reporting, as PPC Land documented, claiming that Gemini 3.5 Flash outperforms the prior-generation Gemini 3.1 Pro on almost all coding benchmarks while running roughly four times faster than other frontier coding models. Google paired that model with Antigravity 2.0, its own agent-first development environment, and with a proposed open web standard called WebMCP intended to let websites expose structured tools directly to AI agents. Whether Muse Spark 1.1's reported gains on Meta's internal coding benchmark would hold up against Gemini 3.5 or other named competitors on a shared, independently run benchmark is not something Meta's announcement addresses, since the company published only its own internal comparison.

The agentic advertising standards conversation forms a third point of context. Industry protocols including the Ad Context Protocol, launched in October 2025 with six founding members, and the Model Context Protocol underpinning it, have divided the advertising industry over questions of fragmentation versus consolidation, as multiple competing frameworks emerged within weeks of each other. A general-purpose reasoning model such as Muse Spark 1.1, capable of generalizing to new MCP servers without additional training, is precisely the kind of component those protocols are designed to plug into. Meta did not name any advertising-specific integration for Muse Spark 1.1 in its July 9 announcement, and none should be assumed from the release itself.

Finally, the framing Kaplan used - describing the release in terms of "US leadership in AI" - situates Muse Spark 1.1 within a wider geopolitical argument Meta has made previously about open and openly-licensed model development. Meta's earlier Llama family of models was positioned in similar terms when the company released Llama 3.1 in 2024, arguing at the time that open-source AI served US and allied interests better than closed alternatives. Muse Spark 1.1's availability is structured differently, however: it reaches outside developers through the Meta Model API rather than as an openly downloadable model, a distinction Meta's July 9 announcement does not explain further.

Timeline

  • July 30, 2025 - Meta establishes Meta Superintelligence Labs, consolidating foundation model, product, and Facebook AI Research teams.
  • October 8, 2025 - Meta begins its phased deprecation of legacy Advantage Shopping and App Campaign APIs, part of the same broader automation push Meta's foundation model work now feeds into.
  • October 15, 2025 - The Ad Context Protocol launches with six founding advertising technology members, built on the Model Context Protocol.
  • January 28, 2026 - Meta reports fourth-quarter 2025 advertising revenue of 58.1 billion US dollars, alongside plans to consolidate more ranking models in 2026 than in the two prior years combined.
  • April 29, 2026 - Meta launches Ads AI Connectors and a companion Ads CLI, giving Claude and ChatGPT write access to campaign creation.
  • May 19, 2026 - Google announces the Gemini 3.5 series and Antigravity 2.0 at Google I/O 2026, a rival long-context, agentic coding model launch.
  • June 30, 2026 - Meta publishes its Developer Tools MCP announcement, a largely read-only server distinct from its write-heavy ads connectors.
  • July 9, 2026 - Meta releases Muse Spark 1.1 and opens a public preview of the Meta Model API to outside developers.

Summary

Who: Meta, through its Meta Superintelligence Labs research division, released the model; Joel Kaplan, Meta's Chief Global Affairs Officer, announced the developer opening publicly, and early assessments came from named executives at Replit, Cline, Box, and the OpenClaw Foundation.

What: Meta released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks with a 1 million token context window, gains in coding and computer-use performance, and a published safety evaluation report, while opening a public preview of the Meta Model API for outside developer access.

When: The model and API preview were announced on July 9, 2026, with Kaplan's public developer announcement following one day later.

Where: The release applies to Meta's global developer platform through the Meta Model API, with the model also available inside the Meta AI app and at meta.ai; Meta did not restrict the announcement to specific countries or regions.

Why: The release matters to the marketing and advertising community because it extends Meta's foundation-model capability at the same time the company is calibrating, tool by tool, how much autonomous access its own advertising infrastructure grants to AI agents - a question with direct consequences for how campaign management, creative production, and account-level automation evolve across the industry in the months ahead.