webAI Releases TwIL-8B Reasoning Model for Math and Code

webAI has released TwIL-8B, a dense 8.79B-parameter reasoning model post-trained from IBM Granite 4.2 8B. It supports English text input and output, uses a 131,072-token context window, and is available under the webAI Non-Commercial 1.0 license. A Q4_K_M GGUF version occupies 4.98 GiB, allowing local inference on a CPU or a GPU with modest VRAM; full bf16 weights require 16.38 GiB.

The model’s strongest reported result is 98.2 on MATH-500. It scores 88.3 on AIME 2026 and 81.6 on LiveCodeBench, although webAI did not record the LiveCodeBench version used. On instruction-following tests, it reaches 74.4 on IFBench and 91.8 on IFEval. Knowledge benchmarks are weaker: GPQA Diamond is 70.2, compared with 81.7 for Qwen3.5-9B, while MMLU-Pro is 76.5.

TwIL-8B has balanced and verbose variants. The former is designed to produce shorter reasoning traces, while the latter targets difficult contest problems. On AIME 2026, webAI reports about 19.5k output tokens per problem, 38% fewer than Qwen3.5-9B, but with a 4.2-point lower score. The published benchmark figures are self-reported and were measured with thinking enabled.