Meta Ships Muse Spark 1.1 as a Closed, Metered API, Abandoning Its Open-Weight Llama Playbook

Meta Superintelligence Labs released Muse Spark 1.1 on July 9, a multimodal reasoning model built for agentic work, alongside a public preview of the Meta Model API (application programming interface). The important detail is not the model but how it is distributed. Unlike the Llama family, whose weights Meta published for download, Muse Spark 1.1 is closed, hosted, and metered per token, reachable only through Meta's apps or the paid API.

The specifications place it directly against the incumbents. The context window is one million tokens, and Meta reports gains over the first Muse Spark in tool use, computer use, coding, and multimodal understanding. Pricing is $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits, and the endpoint is a compatible substitute for the OpenAI and Anthropic software development kits (SDKs). The preview is limited to the United States.

The commercial logic is explicit. A Russian-language technical channel noted that, unlike Llama, the new model's weights are closed, and that the model is designed to operate as the primary agent, assembling context and orchestrating tool calls. Meta is converting its research output into a revenue line, matching the metered-token business that funds OpenAI and Anthropic rather than the open-ecosystem strategy that defined its earlier approach.

The honest read requires caution on the capability claims, which come from Meta's own blog rather than independent reproduction. No third party has yet posted head-to-head numbers on a named agentic benchmark such as SWE-bench or a computer-use suite. What is verified is the pricing, the context length, and the closed licensing, and those are the facts that change how engineers plan.