Meta Launches Muse Spark 1.1 and Opens Meta Model API to Developers

Quick overview: Meta introduced Muse Spark 1.1 on July 9, 2026, a multimodal reasoning model from Meta Superintelligence Labs built for agentic tasks, with major upgrades in tool use, computer use, coding, and multimodal understanding. Alongside the model, Meta opened a public preview of the new Meta Model API, giving outside developers direct access to Muse Spark 1.1 for the first time through an OpenAI compatible interface. The model is also live now in Thinking mode inside the Meta AI app and on meta.ai.

Meta has pulled back the curtain on its next flagship reasoning model, and this time it is not keeping the access limited to internal teams. Muse Spark 1.1 arrives as a direct successor to the original Muse Spark, and Meta is positioning it as a complete agentic foundation rather than a narrow chat model, one built to plan, delegate, browse, code, and operate software largely on its own.

What makes this release notable is not just the model itself but how Meta is choosing to ship it. For the first time, external developers can build on Muse Spark 1.1 through the newly opened Meta Model API, now in public preview. Early partners including Replit, Cline, and Box have already tested the model against real workloads, and their reactions, along with a fresh set of safety disclosures from Meta, give a clearer picture of what this launch actually means for builders.

What Muse Spark 1.1 Brings to the Table

Meta describes Muse Spark 1.1 as advancing what it calls the performance efficiency frontier, meaning it aims to do more per unit of compute than its predecessor. The model ships alongside this week's separate launch of Muse Image, and Meta frames both releases as steps toward its stated goal of personal superintelligence, models meant to help people pursue goals, create, and take action on their behalf.

Four capability areas stand out in Meta's own breakdown.

Agentic orchestration

Muse Spark 1.1 is trained to plan and coordinate work across external apps, native tools, MCP servers, and custom skills without needing task specific fine tuning first. It can act as a lead agent, gathering context and delegating pieces of a task to parallel subagents, or as a subagent itself, staying within its assigned job and escalating back up when needed. Meta also says the model can actively manage a context window of one million tokens, retrieving details from much earlier in a session and compacting history in a way that preserves the steps that matter for later work.

Computer use

The model is built to operate real interfaces across multiple applications as information changes mid task. Rather than clicking through every step individually, Meta says it was trained to judge when writing a script is faster than manual interaction, and when to batch multiple actions together at once. A dinner party planning demo included in the announcement shows the model adjusting an order automatically when new context appears mid task.

Coding

Meta reports substantial gains on real world coding tasks involving large, complex codebases, including bug diagnosis, feature implementation in enterprise systems, and large scale code migrations. The model is designed to work smoothly inside popular agentic coding setups, supporting planning mode, goal conditioning, subagent delegation, and context compaction. On Meta's internal coding benchmark, the company says Muse Spark 1.1 marks a significant jump over the original Muse Spark and is competitive with other leading models on the market.

Multimodal reasoning

The model can inspect images, video, and audio together with text, and Meta highlights strengths in turning visuals into code, producing highly detailed image and video captions, and running agentic workflows that combine perception with action. A demonstration shows the model reviewing smartphone video of a product and then independently creating a Facebook Marketplace listing based on what it observed.

A New Path for Developers: The Meta Model API

Until now, access to Meta's most capable reasoning models has mostly run through the Meta AI app and meta.ai directly. That changes with this release. The new Meta Model API opens a public preview specifically for Muse Spark 1.1, giving developers a direct, OpenAI compatible way to integrate the model into their own products.

Early partner reactions included in Meta's announcement point to a few recurring themes: the value of the one million token context window, native support for images, video, and PDFs, built in search with citations, structured output, and parallel tool calling, all inside a single model. Replit's CEO Amjad Masad called it a "complete agentic foundation." Cline's CEO Saoud Rizwan pointed to Meta's focus on serious agentic coding, noting strong tool use at a price point that makes real coding workloads viable at scale. Box also shared results from testing the model against its own enterprise evaluation set, reporting capabilities competitive with today's leading frontier models, particularly for structured, procedural workflows across professional services, public sector, and industrial use cases.

Safety Testing Behind the Release

Meta says Muse Spark 1.1 went through extensive safety evaluation before launch under its Advanced AI Scaling Framework, the internal policy that sets evaluation methods, threat models, and deployment thresholds for its most advanced models. Across the three frontier risk categories the framework tracks, chemical and biological risk, cybersecurity, and loss of control, Meta states the model operates within safe margins.

The company also reports stronger resistance to both direct jailbreak attempts and indirect attacks, including prompt injection from untrusted data and adversarial developer prompts, along with lower hallucination rates and reduced sycophancy compared to earlier models. Meta has published a full evaluation report covering the model's safety posture for anyone who wants a deeper look at the methodology.

Why This Release Matters

Muse Spark 1.1 lands at a moment when most major AI labs are racing to prove their models can act, not just answer questions. Meta's emphasis on long context management, multi agent orchestration, and real computer use signals it wants Muse Spark to compete directly with the leading agentic coding and automation models already on the market, while the move to open a public API suggests Meta is serious about building a developer ecosystem around it rather than keeping the model locked inside its own consumer apps.

For now, Muse Spark 1.1 is live in Thinking mode inside the Meta AI app and on meta.ai, with the Meta Model API available in public preview for developers who want to build on it directly. Meta has signaled this is not the end of the roadmap, noting in its announcement that additional, more capable models are already in training.

Source: https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/