webAI open-sources TwIL-8B reasoning model with strong math scores

webAI, an AI lab focused on small local professional models, has released the TwIL-8B reasoning model. Built on IBM Granite-4.2-8B, it is optimized for math, coding, and instruction-following tasks, and is open-sourced under a non-commercial license.

In benchmarks, TwIL-8B scored the best among compared models on two advanced math tests, MATH-500 and AIME 2025, even beating Google's Gemma 4 31B, a model several times its size. It also had the best LiveCodeBench score in its group and matched Gemma 4 12B on the IFBench instruction test.

webAI says TwIL-8B's reasoning reaches 88% to 93% of frontier models in its strong areas, which can cut AI costs in many cases. However, it has limited knowledge and is not a general-purpose model.