Meta’s new Muse Spark 1.1 pushes agentic AI with a million-token window
Meta is quietly shifting how developers access advanced AI. The company’s Superintelligence Labs just released Muse Spark 1.1, a closed, hosted reasoning model built for agentic tasks, and opened the Meta Model API to the public. Unlike Meta’s open-weight releases, Muse Spark 1.1 is metered by token and positioned as a specialized orchestration layer—one that thinks before it answers, calls tools in parallel, and can manage a context window measured in the millions of tokens.
A model that plans, then acts
Muse Spark 1.1 is explicitly a reasoning model. Before producing an output, it performs adjustable internal reasoning steps, a capability Meta says improves accuracy on complex tasks such as tool use, computer interaction, coding, and multimodal understanding. The model accepts text, images, video, and documents, and returns text. It also supports structured outputs, prompt caching, and a Files API. For developers who need grounded answers, a web_search tool can be embedded in the Responses API call to surface cited sources.
Hosted, priced, and US-first for now
Meta is charging $1.25 per million input tokens and $4.25 per million output tokens, with new accounts receiving $20 in free credits. Consumers can try the model for free in “Thinking” mode inside the Meta AI app and on meta.ai, but developers outside the US and EU will have to wait: the public preview is currently US-only.
Benchmarks and positioning
Meta’s internal benchmarks show Muse Spark 1.1 leading on tool-use and tool-augmented reasoning tasks, while ranking third on coding and multimodal reasoning. The company emphasizes orchestration behavior—active context compaction, delegation to subagents, and zero-shot generalization to new tools—as the real differentiator. In practice, that means the model can remember actions across a million-token window, plan a workflow, farm out subtasks to specialized subagents, and escalate when it hits a limit.
Wiring it into an existing stack is intentionally simple: the Meta Model API is OpenAI-compatible, so a base-URL change and an extra header are often all that’s needed. Meta even provides a starter snippet using the standard OpenAI SDK.
Why it matters
Meta is betting that developers will pay for hosted, reasoning-first models that can orchestrate complex workflows rather than just answer single-turn questions. By coupling Muse Spark 1.1 with a developer-friendly API and generous free tier, the company is creating a low-friction path to agentic AI—one that may appeal to teams that lack the infrastructure to host large reasoning models themselves. The million-token context and adjustable reasoning effort hint at a future where AI systems can handle long-running, multi-step tasks without constant human oversight.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.