Summary: Meta released Muse Spark 1.1 on July 9, 2026 and opened the Meta Model API in public preview. The multimodal model is described as handling 1 million-token context compaction, new tools and MCP servers, multi-agent orchestration, coding, and choices between scripts and direct computer use. Partner quotes and safety results are launch-day claims, while pricing, rate limits, and migration details are absent.
Anyone who ships agents learns quickly that the model call is not where the pain lives. The trouble comes from context that quietly drifts, tools that fall over the moment they meet an interface they were never tested against, and long sessions in which the model loses track of work it already finished. Meta’s Muse Spark 1.1, released on July 9, 2026, aims straight at those failure modes, and it arrives together with a public preview of the Meta Model API.
What the release actually claims
Meta’s announcement describes Muse Spark 1.1 as a multimodal reasoning model built for agentic tasks, with gains over the earlier Muse Spark in tool use, computer use, coding, and multimodal understanding. You can use it now in “Thinking” mode inside the Meta AI app and on meta.ai; for developers, the way in is the new Meta Model API, currently in public preview.
The agentic claims are concrete rather than vague. Meta says the model zero-shot generalizes to new native tools, MCP servers, and custom skills, and that it has been trained to orchestrate multi-agent systems with end-to-end latency as the goal. In the main-agent role it collects context, plans, and hands execution off to parallel subagents; dropped into a subagent role, it holds its assignment and recognizes the moment to escalate. It also manages a 1 million token context window, recalling earlier actions and compacting as it goes while preserving the steps that later work depends on.
Computer use and coding: the tradeoff Meta is naming
The computer-use section spells out a design decision that deserves a second read. Instead of reasoning through every desktop step one click at a time, Meta says, the model judges when a script is the better tool and when clicking directly wins, emitting batches of actions per step. That is a genuine tradeoff: a script runs faster and repeats more reliably, while direct interaction is what carries you through interfaces that resist automation.
On coding, Meta reports substantial improvement on large, complex codebases, including diagnosing bugs, building features inside enterprise systems, and carrying out large migrations. The announcement says the model adapts to diverse harnesses and handles multi-turn dynamics, with support for planning mode, goal conditioning, subagent delegation, and context compaction. Meta points to its internal coding benchmark, Meta Internal Coding Bench, as significantly improved over Muse Spark and competitive with leading alternatives.
Partner signals, with the usual caveat
The announcement carries quotes from Replit, Cline, Box, and the OpenClaw Foundation. Amjad Masad of Replit describes a million-token context, multimodal support, built-in search with citations, structured output, and parallel tool calling, all wrapped in an OpenAI-compatible package. Saoud Rizwan of Cline points to tool use at a price point that makes running real coding workloads at scale viable. Bear in mind that these are launch partners speaking on launch day; read them as directional signals, not as independent benchmarks.
Safety posture and what is not specified
Meta says it ran safety evaluations under its Advanced AI Scaling Framework covering the Chemical & Biological, Cybersecurity, and Loss of Control categories, and reports that the model operates within safe margins, with stronger adversarial robustness, lower hallucination rates, and less sycophancy. The complete posture lives in the Muse Spark 1.1 Evaluation Report.
What builders do not get from the announcement: pricing, rate limits, or a migration path away from other providers. If you are weighing this against the stack you run today, those are the questions worth answering before you commit.
How to evaluate it without wasting a sprint
For teams already running agents in production, the practical move is to test the two claims that change architecture: how context compaction behaves over long sessions, and whether the model really switches between scripting and direct UI interaction inside your own workflows. That second behavior is exactly the kind that shines in a demo and then fails quietly the first time it meets an unfamiliar internal tool.
Teams deciding where agent execution should live will find a useful companion in the earlier post on what a hosted sandboxed shell changes for agentic tools, because tool execution and context management tend to break at the same time. Pick one workflow you already measure, and see whether compaction still holds past the point where your current model begins dropping the thread.