webAI releases TwIL-8B, an 8.8B reasoning model scoring 98.2 on MATH-500

The model, post-trained from IBM Granite 4.2 8B, posts strong math and coding results but trails peers on knowledge benchmarks.