Meta deployed its latest artificial intelligence model via the Meta Model API and Muse Code on Sept. 2 to increase agentic and coding capabilities. The model achieved a score of 88.8 on Terminal-Bench 2.1, matching GPT-5.6 Sol, and recorded 75.4 on DeepSWE. Internal comparisons show the update utilizes 20% fewer tool calls and 25% fewer tokens than the Muse Spark 1.2 version.
The update sustains long-horizon work across multiple workflows in a single thread and is calibrated to avoid hallucinating outcomes. CEO Mark Zuckerberg announced that a max reasoning mode will be released after safety testing, followed by a model named Watermelon and open weights for the Muse Spark line. The model is currently accessible through the Meta Model API, Muse Code, and OpenRouter.
Key sources
- SOURCE@finkd“frontier performance almost too cheap to meter”x.com
- SUPPORT@alexandr_wang“holds onto requirements well during long-horizon tasks”x.com
- SUPPORT@aiatmeta“asks clarifying questions, flags when it's stuck, confirms before consequential actions”x.com
- SUPPORT@aiatmeta“max reasoning coming soon after we finish safety testing”x.com
- SUPPORT@wallstengine“scoring 75.4 on DeepSWE”x.com
- SUPPORT@levie“completely changes the dynamic of US open weights competitiveness”x.com
- SUPPORT@business“capabilities are edging closer to top competitors”x.com