webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500

5 November 2026 4 min read هدهد تك | Hudhud Tech

webAI has released TwIL-8B, an 8.8B reasoning model designed for math, coding, and instruction following. This model is post-trained from IBM’s Granite 4.2 8B. The model’s weights are available on Hugging Face under a non-commercial license. According to webAI’s launch post, the focus is on cost per task rather than raw scale. The 4-bit version of the model is 4.98 GiB, making it suitable for CPU or modest VRAM GPU inference.

TL;DR

Size: 8.79B parameters, dense, with a 131,072-token context inherited from the base model.
Runs on: CPU or modest VRAM GPU via a 4.98 GiB Q4_K_M GGUF; full bf16 weights are 16.38 GiB.
Performance: Leads its size class in math and coding but falls short in knowledge benchmarks.
Best: MATH-500 score of 98.2, the highest in webAI’s comparison.
Worst: GPQA Diamond score of 70.2, trailing Qwen3.5-9B by 11.5 points.

What is TwIL-8B?

TwIL-8B is a dense decoder-only reasoning model from webAI. It features 40 layers, a hidden size of 4096, and 32 attention heads with 8 KV heads. The model processes input and output exclusively in English text and includes a <think> block before each answer.

It is available in two variants with identical architecture and tokenizer: the Balanced variant, which produces shorter reasoning traces, and the Verbose variant, which generates longer reasoning suitable for challenging contest-style problems.

TwIL-8B is part of webAI’s TwIL family, which began with the release of 1.7B and 3B formal-logic models in August 2026.

How was TwIL-8B trained?

webAI utilized three post-training stages:

Full-parameter SFT on reasoning traces for math, STEM, and instruction following.
GRPO/DAPO reinforcement learning on math, coding, STEM, and instruction prompts, producing the verbose variant.
Length-aware RL with a reward system favoring shorter traces compared to per-domain length references, resulting in the balanced variant.

The training datasets were decontaminated using 13-gram matching against AIME, HMMT, MATH-500, GPQA, and other evaluations.

How does TwIL-8B perform on benchmarks?

All scores for TwIL-8B are self-reported and reflect high-effort pass@1 with thinking enabled. Scores marked with a † were measured on the verbose variant only.

Math: TwIL-8B achieves a score of 98.2 on MATH-500, while Qwen3.5-9B scores 96.6 on a third-party leaderboard. Gemma 4 31B scored 87.9 in Red Hat’s lighteval run. On AIME 2025, TwIL-8B scores 88.3, compared to 87.1 for Gemma 4 31B, and on AIME 2026, it scores 88.3, trailing Qwen3.5-9B at 92.5 and Gemma 4 31B at 89.2.
Code: LiveCodeBench† score is 81.6, compared to 65.6 for Qwen3.5-9B and 80.0 for Gemma 4 31B. webAI notes that its LiveCodeBench version was not recorded; competitor scores are from v6.
Instruction following: IFBench† score is 74.4, which is 9.9 points higher than Qwen3.5-9B at 64.5. IFEval results are similar, with TwIL-8B at 91.8 versus 91.5 for Qwen3.5-9B.
Knowledge: This area shows weaker performance, with a GPQA Diamond score of 70.2, compared to 81.7 for Qwen3.5-9B and 84.3 for Gemma 4 31B. MMLU-Pro† score is 76.5 against 82.5 for Qwen3.5-9B.

Across 13 shared benchmarks, TwIL-8B has a record of 7-6 against Qwen3.5-9B and 3-11 against Gemma 4 31B, which is approximately 3.5 times its size.

Does TwIL-8B use fewer reasoning tokens?

Yes, compared to its closest open peer. On AIME 2026, TwIL-8B averages about 19.5k output tokens per problem, while Qwen3.5-9B uses 31.2k, resulting in a 38% reduction in token usage. However, TwIL-8B scores 4.2 points lower. On HMMT February 2026, the difference is 32.9k versus 40.6k, or 19% fewer tokens.

Frontier models are significantly leaner, with webAI citing around 5k tokens per AIME 2026 problem for GPT-5.5. webAI notes that harnesses and tokenizers differ, so these figures should be treated as directional.

TwIL-8B vs Qwen3.5-9B vs Gemma 4

TwIL-8B figures from the model card. Scores marked with † are from the verbose variant. All scores are self-reported by each vendor unless otherwise noted.