Meta’s Muse Spark 1.3 lands on Google Cloud with 1M context

Meta’s Muse Spark 1.3 is in public preview on Google Cloud alongside Muse Wonder and Muse Ember. Google Cloud’s model page, updated November 9, gives it a 1,048,576-token input limit. It also lists batch predictions and Provisioned Throughput as not supported, so Spark 1.3 runs only on standard pay-as-you-go or a fixed quota, and only through the global endpoint.

The page gives two forms of the model ID. The catalog lists it as meta/muse-spark-1.3, but requests to the global Responses endpoint, /locations/global/publishers/meta/v1/responses, pass the bare muse-spark-1.3 in the model field.

Google Cloud describes Spark 1.3 as a reasoning model trained for agentic workflows and competitive coding. It has built-in tool calling with support for MCP, the Model Context Protocol that lets a model reach external tools and data through one interface. It takes text, images and PDFs and returns text. The full million-token window is aimed at multi-step tasks, and the quota for the global endpoint allows the same context length.

The long window is not new for the Muse line, which Meta Superintelligence Labs has built since April as the successor to Llama. Meta’s own Model API opened in public preview with Muse Spark 1.1 on July 9, already with a million-token context, and widened to developers worldwide on August 5 with Spark 1.2.

The page dates this version September 24. Function calling, structured output and reasoning each still carry a Preview tag there.