webAI open-sources TwIL-8B reasoning model, beating Gemma 4 31B on math and code

webAI, an AI lab focused on small specialized models for local use, launched the TwIL-8B reasoning model on November 4, 2026. The launch was covered on November 6 by IT之家. The model is built on IBM Granite-4.2-8B and is optimized for math, coding, and instruction-following workloads. It is open-sourced under a non-commercial license, with the model card published on Hugging Face at huggingface.co/webAI-Official/TwIL-8B.

According to webAI, TwIL-8B achieves the best scores among the models it compared on MATH-500 and AIME 2025, two higher-level math benchmarks, outperforming Google's Gemma 4 31B despite being several times smaller. Its LiveCodeBench result is also reported as the best in its group, and its IFBench instruction-execution performance is said to match Gemma 4 12B. webAI further claims the model's reasoning ability reaches 88~93% of frontier models on its strengths, which could lower AI usage costs in many cases. The lab also cautions that the model has limited knowledge and is not a general-purpose one.

So far only webAI's own statements describe TwIL-8B; the reported results have not been independently verified, and no further developments have been announced.