Thinking Machines Unveils Inkling 2 Small for Enterprise AI Workloads

Thinking Machines Lab has officially introduced Inkling 2 Small, a powerful new artificial intelligence model designed specifically for complex enterprise workflows and sustained agentic tasks. The November 2026 release marks a significant architectural upgrade within the company model family. It features a highly efficient Mixture of Experts design containing 594 billion total parameters while activating only 17 billion parameters per token.

This sparse activation strategy allows the system to handle demanding reasoning tasks, advanced software development, and detailed visual document analysis without overwhelming corporate computational resources. According to the official launch announcement, the model natively supports both text and image inputs, making it a highly versatile tool for modern multimodal business applications.

The new model demonstrates impressive capabilities across multiple independent evaluation platforms. It achieved a remarkable score of 31 on the Artificial Analysis Intelligence Index, positioning it as a top performing open weight alternative in the Western market. For corporate environments, the system scored 71.8 percent on the EnterpriseBench CoreCraft benchmark and 46.8 percent on EnterpriseOps Gym. These specific metrics prove its ability to manage simulated customer service interactions and complex IT operations effectively. The results indicate that the architecture is well suited for professional roles requiring repeated tool calls, iterative code revisions, and precise financial document analysis.

Organizations can currently access the model through the proprietary Tinker platform for both inference and custom fine tuning. The pricing structure is set at 30 cents per million input tokens and 1 dollar per million output tokens, offering a competitive entry point for developers testing autonomous workflows. Thinking Machines has firmly committed to releasing the full model weights under an Apache 2.0 license following the completion of rigorous external safety testing and red teaming protocols. This upcoming open release will allow enterprises to self host the technology, ensuring strict data privacy and complete operational control over customized deployments.

While the performance metrics are undeniably strong, technical teams must account for specific operational constraints before full integration. The current iteration does not process audio inputs, requiring separate transcription services for any voice based applications. Additionally, the model can be quite verbose during complex reasoning tasks, which may inadvertently increase token consumption and overall operating costs.

Enterprises are strongly advised to implement strict network restrictions and granular tool permissions when deploying autonomous agents. By treating this technology as a customizable assistant rather than a fully independent operator, businesses can safely leverage its advanced capabilities for rapid prototyping, financial analysis, and secure software development.