Inkling 2 Small: Thinking Machines introduces new Mixture-of-Experts model

Thinking Machines has officially introduced its second-generation AI model, Inkling 2 Small. Built on a Mixture-of-Experts architecture, the new system features 594 billion total parameters with 17 billion active parameters, alongside support for image inputs and variable reasoning effort.

• Thinking Machines unveiled Inkling 2 Small, a new Mixture-of-Experts AI model.
• The model features 594 billion total parameters and 17 billion active parameters.
• It supports multimodal inputs (images) and excels in agentic coding tasks.
• Full model weights under the Apache 2.0 license will be released next week following external safety testing.

Architecture and Performance Improvements Inkling 2 Small significantly outperforms the company's previous models on the Artificial Analysis Intelligence Index with a score of 31. This performance boost stems from improvements made throughout the training process, including more efficient pre-training and a large-scale reinforcement learning (RL) run. The model shows major gains in agentic coding, computer vision, knowledge work, and complex reasoning. Designed for long-horizon autonomous work, Inkling 2 Small was trained on hundreds of thousands of coding tasks using advanced RL strategies. Unlike simpler models, it evaluates problem-solving approaches rather than relying solely on unit-test rewards, making it exceptionally skilled at following instructions and self-correcting during real-world agent loops.

Vision Capabilities and Technical Tasks The model incorporates a 292-million-parameter vision encoder based on Perception Encoder. This enables powerful visual capabilities across categories such as optical character recognition (OCR), infographic interpretation, and visual reasoning. In a notable demonstration, the model successfully identified wine varieties and vintages from a single photograph during a blind tasting test. Furthermore, Inkling 2 Small is capable of handling long-horizon autoresearch tasks, including GPU kernel optimization and machine learning experimentation. It demonstrated the ability to autonomously train models, write its own evaluators, and iterate on training recipes to improve forecasting scores over several hours of execution.

Availability and Open Weights Roadmap Initially, Inkling 2 Small is available through the Tinker platform with closed weights as part of the company's safe release path. Users can test and fine-tune the model directly within Tinker Playground. The creators intend to publish the full model weights under an Apache 2.0 license next week, once external safety testing and red-teaming procedures are successfully completed.