Meta Muse Spark 1.1: the agent race gets a third serious model
The same day OpenAI shipped ChatGPT Work, Meta Superintelligence Labs dropped Muse Spark 1.1 and opened the Meta Model API in public preview. This is not a chat model with tools tacked on β it is built for agents from the ground up. Multimodal reasoning, a live one-million-token context it actively manages, and the ability to orchestrate parallel subagents on a complex project: gather context, plan, delegate, keep going.
The capability list reads like a checklist of everything agents actually break on. Computer use across multiple apps in extended sessions. Coding that diagnoses bugs and runs large migrations, with real support for planning mode and subagent delegation. Visual-to-code, video and image captioning, multimodal workflow execution. And the context-management piece matters more than it sounds β remembering what it did and compacting strategically is exactly where long agent runs fall apart today.
Step back and the picture is a three-way lab race for the agentic foundation model, all breaking cover the same week. OpenAI has GPT-5.6 under ChatGPT Work. Anthropic has the Claude line under its harness. Now Meta wants developers building on the Meta Model API, with Muse Spark also live in Thinking mode inside the Meta AI app. Meta spent a year rebuilding around Superintelligence Labs, and this is the first shot that looks like it belongs in the same room.
The open question is distribution. OpenAI and Anthropic already own the developer mindshare and the agent tooling; a strong model is table stakes, not a win. But a million-token context with full multimodal support and genuine subagent orchestration is a complete agentic base, and Meta pricing it as an open API is the aggressive move. Details at https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/
← Back to all articles
The capability list reads like a checklist of everything agents actually break on. Computer use across multiple apps in extended sessions. Coding that diagnoses bugs and runs large migrations, with real support for planning mode and subagent delegation. Visual-to-code, video and image captioning, multimodal workflow execution. And the context-management piece matters more than it sounds β remembering what it did and compacting strategically is exactly where long agent runs fall apart today.
Step back and the picture is a three-way lab race for the agentic foundation model, all breaking cover the same week. OpenAI has GPT-5.6 under ChatGPT Work. Anthropic has the Claude line under its harness. Now Meta wants developers building on the Meta Model API, with Muse Spark also live in Thinking mode inside the Meta AI app. Meta spent a year rebuilding around Superintelligence Labs, and this is the first shot that looks like it belongs in the same room.
The open question is distribution. OpenAI and Anthropic already own the developer mindshare and the agent tooling; a strong model is table stakes, not a win. But a million-token context with full multimodal support and genuine subagent orchestration is a complete agentic base, and Meta pricing it as an open API is the aggressive move. Details at https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/
Comments