Thinking Machines, a startup founded by former OpenAI CTO Mira Murati, has introduced Inkling 2 Small. This mixture-of-experts model features 594 billion total parameters, with 17 billion active at any given time. It supports both text and image inputs and allows users to adjust thinking depth to balance speed and thoroughness.
The model shows significant performance improvements over its predecessor, Inkling Small. On the Artificial Analysis Intelligence Index, it scored 31 points, compared to 13 for the previous version. It also achieved a score of 86.3 percent on SWEBench Pro v2 (Hard) and 81.6 percent on the MMMU Pro benchmark for multimodal academic tasks.
Operational costs for Inkling 2 Small are $1.18 per task, which is nearly six times higher than the predecessor. The model consumes approximately 150,000 output tokens per task. Despite the cost increase, it remains significantly cheaper than Gemini 3.8 Flash at $7.77 and Claude Sonnet 5.5 at $10.02 per task.
Training involved processing 12 trillion tokens in total. The reinforcement-learning phase utilized 2,048 GPUs over a nine-day period, consuming 3 trillion tokens. The company plans to release the model weights under the Apache 2.0 license following the completion of external safety testing.
The model currently demonstrates strengths in visual reasoning, infographic interpretation, and text recognition. It leads among open-weight models on the DesignArena Website benchmark with an Elo rating of 1,340. However, the company notes limitations, including inefficient communication and a lack of audio support.
