If you build agents, the hard part is rarely the model call. It is context that drifts, tools that break on unfamiliar interfaces, and long sessions where the model forgets what it already did. Meta’s July 9, 2026 release of Muse Spark 1.1 targets exactly those failure modes, and it ships alongside a public preview of the Meta Model API.
What the release actually claims
According to Meta’s announcement, Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with gains in tool use, computer use, coding, and multimodal understanding over the earlier Muse Spark. The model is available now in “Thinking” mode in the Meta AI app and on meta.ai, and developers can access it through the new Meta Model API in public preview.
The agentic claims are specific. Meta says the model zero-shot generalizes to new native tools, MCP servers, and custom skills, and that it is trained to orchestrate multi-agent systems for end-to-end latency. As a main agent it gathers context, plans, and delegates to parallel subagents; as a subagent it stays on task and knows when to escalate. It also manages a 1 million token context window, remembering earlier actions and compacting while keeping critical steps.
Computer use and coding: the tradeoff Meta is naming
The computer-use section describes a design choice worth reading twice. Rather than reasoning through every desktop step one click at a time, Meta says the model decides when to write a script and when to click directly, generating batches of actions per step. That is a real tradeoff: scripting is faster and more repeatable, but direct interaction handles interfaces that resist automation.
On coding, Meta reports substantial improvement on large, complex codebases, including bug diagnosis, feature work in enterprise systems, and large migrations. The announcement says the model adapts to diverse harnesses and handles multi-turn dynamics, supporting planning mode, goal conditioning, subagent delegation, and context compaction. Meta’s internal coding benchmark, Meta Internal Coding Bench, is cited as significantly improved over Muse Spark and competitive with leading alternatives.
Partner signals, with the usual caveat
The announcement includes 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 in an OpenAI-compatible package. Saoud Rizwan of Cline points to tool use at a price point viable for running real coding workloads at scale. These are launch partners speaking at launch; treat them as directional, not as independent benchmarks.
Safety posture and what is not specified
Meta says it ran safety evaluations under its Advanced AI Scaling Framework across Chemical & Biological, Cybersecurity, and Loss of Control categories, and reports the model operates within safe margins, with better adversarial robustness, lower hallucination rates, and reduced sycophancy. The full posture is in the Muse Spark 1.1 Evaluation Report.
What the announcement does not give builders: pricing, rate limits, or a migration path from other providers. If you are weighing this against your current stack, those are the questions to ask before committing.
How to evaluate it without wasting a sprint
If you already run agents in production, the practical move is to test the two claims that change architecture: context compaction over long sessions, and whether the model actually switches between scripting and direct UI interaction on your own workflows. That second behavior is the kind of thing that looks great in a demo and fails quietly on an unfamiliar internal tool.
For teams deciding where agent execution should live, the earlier post on what a hosted sandboxed shell changes for agentic tools is a useful companion, since tool execution and context management tend to fail together. Start with one workflow you already measure, and check whether the compaction behavior holds up past the point where your current model starts losing the thread.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.