Thinking Machines has released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model that claims the top spot among western open-weight models. Founded by former OpenAI CTO Mira Murati, the startup designed the model to accept text and images while offering an adjustable thinking depth for speed or thoroughness. On the Artificial Analysis Intelligence Index, the model scores 31, marking an 18-point improvement over its predecessor and matching Google's Gemini 3.8 Flash.
The model demonstrates significant gains in agentic tasks and coding, jumping from 5 to 46 percent on AutomationBench-AA and reaching 86.3 percent on SWEBench Pro v2. Thinking Machines trained the system using reinforcement learning on hundreds of thousands of long-horizon coding tasks. In addition, the model excels in visual understanding and design, achieving the highest scores among open-weight models on the GDP.pdf document benchmark and the DesignArena Website visual creation test.
These performance improvements come with increased computational costs, generating more output tokens and costing nearly six times as much per task as the previous generation. The model is currently available as a proprietary offering via the Tinker platform, with weights under an Apache license expected to follow. This release matters because it establishes a powerful new open-weight competitor in the AI landscape with strong capabilities in multimodal reasoning and agentic workflows.