Thinking Machines launches Inkling 2 Small for autonomous work

Thinking Machines Lab introduced Inkling 2 Small on November 5, its second-generation model for coding, research and tool use. “Small” describes its active computation: the mixture-of-experts model has 594 billion parameters, with 17 billion active. Its model card lists text and image inputs, text outputs and a context window of up to one million tokens.

The company reports 25.4% on FrontierCode v2.0 and 81.6% on MMMU Pro, alongside an Artificial Analysis Intelligence Index score rounded to 31. Its demos show agents using visual feedback to refine software, including a Three.js city simulator with procedurally generated buildings and cars. These are launch results reported by the developer.

City-simulator demo. Thinking Machines Lab.

One research example is more concrete than a benchmark: an agent trained Qwen3.5-4B to forecast time series, spending $426 on Tinker. The company says its final checkpoint made about 15% less error than the base model on unseen data. It also acknowledges weaknesses in tool selection and chat presentation.

Forecasting experiment and evaluation replay. Thinking Machines Lab.

The model launched on Tinker for inference and fine-tuning with closed weights. Thinking Machines planned an Apache 2.0 weight release the following week, conditional on external safety testing and red-teaming; that plan is not confirmation the weights have shipped.

Sources: Announcement, Model card, Tinker Playground