webAI unveils TwIL-8B, an 8.8B reasoning model achieving high scores on math-500

webAI has introduced TwIL-8B, an 8.8 billion parameter reasoning model designed for tasks involving mathematics, code, and instruction following. This dense, decoder-only model, which is post-trained from IBM’s Granite 4.2 8B, features 40 layers and is available under a non-commercial license on Hugging Face. The model is offered in two variants—Balanced and Verbose—with the latter targeting complex, contest-style problems. The 4-bit quantized version of TwIL-8B is compact at 4.98 GiB, enabling inference on modest hardware. In performance comparisons, TwIL-8B achieves a 7-6 advantage over Qwen3.5-9B across 13 shared benchmarks, though it trails Gemma 4 31B by a 3-11 margin. On the AIME 2026 benchmark, TwIL-8B uses 38% fewer output tokens (19.5k vs 31.2k) than Qwen3.5-9B. webAI noted that the model's scores are self-reported, based on high-effort pass@1 with thinking enabled. The model extends webAI’s TwIL family, which previously included 1.7B and 3B formal-logic models released in August 2026. The company also pointed out that GPT-5.5 reportedly requires about 5k tokens per AIME 2026 problem, suggesting frontier models remain significantly leaner.