Thinking Machines Lab Releases Inkling 2 Small, Pushing Open-Weight Multimodal AI for Customization

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has unveiled its second-generation multimodal AI model, Inkling 2 Small. This new model, introduced on November 5th, is a Mixture-of-Experts (MoE) architecture with 594 billion total parameters and 17 billion active. It supports image inputs and variable effort, and its weights are slated for release next week under an Apache 2.0 license, following external safety testing. The significance for practitioners lies in the model's open-weight nature and its enhanced capabilities. Inkling 2 Small demonstrates notable improvements in areas critical for real-world applications, including agentic coding, vision, knowledge work, and reasoning. This is attributed to more efficient pre-training and a large, unified reinforcement-learning run. The model scores 31 on the Artificial Analysis Intelligence Index, a substantial gain over its predecessors. This release fits squarely within the broader, well-established trend of democratizing AI development and fostering greater control over AI systems. The shift towards open-weight models empowers organizations to move beyond reliance on proprietary, black-box solutions. This allows for greater transparency, auditability, and the crucial ability to fine-tune models on sensitive, domain-specific datasets without relinquishing control or intellectual property. The availability of platforms like Tinker, which provide frontier-grade infrastructure for training custom models, further accelerates this trend. This echoes the sentiment that the AI market is maturing from a focus on "can we make this multimodal?" to "how do we ship a multimodal agent that does not break in production?" In practice, this means that DevOps and AI teams can leverage Inkling 2 Small as a robust foundation for building highly specialized multimodal applications. Organizations can train the model on their unique data, ensuring that the AI aligns with their specific operational needs and compliance requirements. This is particularly valuable in sectors where data privacy and proprietary knowledge are paramount. Practitioners should closely watch the open-weight release and explore the Tinker platform for fine-tuning opportunities. The ability to customize and deploy such a capable multimodal model on their own terms offers a significant competitive advantage, enabling more efficient workflows and the development of innovative AI-powered solutions. The ongoing competition in the open-weight model space, with other significant releases from Reflection AI, Mistral AI, and Tencent, indicates a vibrant ecosystem where choice and customization will be key for adoption.