Meta unveiled "Muse Spark 1.1" on July 9, 2026, a multimodal inference model optimized for agent-based tasks. The model enhances capabilities in tool usage, computer operation, coding, and media understanding. It is currently available within the Thinking mode of the Meta AI application and the web version of meta.ai. Developed by Meta Superintelligence Labs (MSL), the model also marks the launch of a public preview for the Meta Model API, currently limited to the United States.

Muse Spark 1.1 features a context window of up to 1 million tokens, allowing the model to compress context and reference past actions during prolonged operations. API pricing is set at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens, with new accounts receiving $20 in free credits. The API uses an OpenAI-compatible format and supports structured output and parallel tool calling.

The system employs a core agent that develops plans and distributes tasks among multiple sub-agents. It integrates with external tools and MCP servers without requiring individual fine-tuning. In performance testing, the model scored 72.2 on Vibe Code Bench v1.1 for web application creation and 42.0 on SWE Atlas for codebase Q&A. Its JobBench score, which measures practical task performance, rose from 17.0 in the previous version to 54.7.

Meta previously released the "Muse Image" and "Muse Video" models on July 7, 2026. Integrating these with Muse Spark 1.1 enables applications like full-stack website and game production. Security evaluations indicate potential risks in chemical, biological, and cybersecurity domains when safety measures are disabled, leading the company to recommend that developers implement isolated execution environments. Availability of the API in Japan remains undetermined.

Meta unveiled "Muse Spark 1.1" on July 9, 2026, a multimodal inference model optimized for agent-based tasks. The model enhances capabilities in tool usage, computer operation, coding, and media understanding. It is currently available within the Thinking mode of the Meta AI application and the web version of meta.ai. Developed by Meta Superintelligence Labs (MSL), the model also marks the launch of a public preview for the Meta Model API, currently limited to the United States.

Muse Spark 1.1 features a context window of up to 1 million tokens, allowing the model to compress context and reference past actions during prolonged operations. API pricing is set at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens, with new accounts receiving $20 in free credits. The API uses an OpenAI-compatible format and supports structured output and parallel tool calling.

The system employs a core agent that develops plans and distributes tasks among multiple sub-agents. It integrates with external tools and MCP servers without requiring individual fine-tuning. In performance testing, the model scored 72.2 on Vibe Code Bench v1.1 for web application creation and 42.0 on SWE Atlas for codebase Q&A. Its JobBench score, which measures practical task performance, rose from 17.0 in the previous version to 54.7.

Meta previously released the "Muse Image" and "Muse Video" models on July 7, 2026. Integrating these with Muse Spark 1.1 enables applications like full-stack website and game production. Security evaluations indicate potential risks in chemical, biological, and cybersecurity domains when safety measures are disabled, leading the company to recommend that developers implement isolated execution environments. Availability of the API in Japan remains undetermined.