Meta Muse Spark 1.1 Opens Model API Access, Highlighting One-Million-Token Context, Parallel Tool Calling, and Multimodal Agents

Meta is moving Muse Spark 1.1 from a research showcase to a public preview that developers can access through the new Meta Model API. Officially listed capabilities include a one-million-token context window; image, video, and PDF inputs; search with source citations; structured outputs; and parallel tool calling. The API uses an OpenAI-compatible format, aiming to let existing agents switch models with fewer interface changes. This is the most significant engineering aspect of Meta’s latest update: beyond model quality, tools that can actually be orchestrated and response formats are finally becoming part of a public product.

Parallel tool calling is well suited to running independent searches, database queries, or code analyses simultaneously, theoretically reducing wall-clock time for multi-step agents. Meanwhile, the one-million-token context window can accommodate large codebases, document collections, or long-term work logs. However, the cost and attention quality of long context do not necessarily scale linearly. Putting all available data directly into a prompt can still dilute critical information. Search citations improve source visibility, but they do not guarantee source quality or correct reasoning.

Meta has published positive comments from partners about coding and enterprise workflows, but these are not controlled, independent benchmarks. Engineers should evaluate the model on real-world agent tasks, comparing tool-selection accuracy, dependency handling for parallel calls, JSON Schema compliance rates, retry behavior, caching, and pricing for one million tokens. A public preview also means that quotas, regional availability, version stability, and data-processing terms may change. Even with interface compatibility, teams should retain model-routing and fallback capabilities.