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

The 8.8B model, post-trained from IBM Granite 4.2 8B, scores well on math and coding benchmarks but trails peers on knowledge tasks.