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

The 8.8B model is post-trained from IBM Granite 4.2 8B and trails comparable models on knowledge benchmarks.