webAI Releases TwIL-8B Reasoning Model Built on IBM's Granite for Local Use

The release joins a fast-growing group of small open-weight models that companies can run on their own machines instead of paying per token for frontier models in the cloud. webAI is pitching the model on cost per task rather than raw scale. A 4-bit quantised build is 4.98 GiB, small enough for a CPU or a GPU with modest memory, while the full bf16 weights take 16.38 GiB.

The licence is the main catch for businesses. IBM's Granite base model is released under Apache 2.0, but TwIL-8B carries the "webAI Non-Commercial License ver. 1.0", so companies cannot use it in commercial products under the published terms.

Strong on Maths, Weak on Knowledge

On webAI's model card, TwIL-8B scores 98.2 on MATH-500. That compares with 96.6 for Alibaba's Qwen3.5-9B on a third-party leaderboard and 87.9 for Google's Gemma 4 31B in a Red Hat evaluation. On LiveCodeBench, its longer-reasoning variant scores 81.6, against 80.0 for Gemma 4 31B and 65.6 for Qwen3.5-9B. webAI notes that it did not record which version of LiveCodeBench it used, while the competitor scores are on version 6.

The results are mixed elsewhere. On AIME 2026 the model scores 88.3, behind Qwen3.5-9B at 92.5 and Gemma 4 31B at 89.2. It is weakest on knowledge-heavy tests. It scores 70.2 on GPQA Diamond, against 81.7 for Qwen3.5-9B and 84.3 for Gemma 4 31B. Across 13 benchmarks both models report, TwIL-8B beats Qwen3.5-9B on seven, according to a MarkTechPost tally. Against Gemma 4 31B, a model about 3.5 times its size, it wins three and loses 11. All of webAI's scores are self-reported.

Shorter Reasoning Traces

The model comes in two versions with the same architecture. The default "balanced" version produces shorter chains of reasoning. The "verbose" version reasons for longer and is meant for hard, contest-style problems where accuracy matters more than cost.

webAI trained the model in three stages: supervised fine-tuning on reasoning traces, then reinforcement learning with the GRPO and DAPO methods, which produced the verbose version. A final reinforcement learning stage rewarded shorter answers to produce the balanced version. The company says it checked its training data against AIME, HMMT, MATH-500, GPQA and other test sets to remove overlap.

On AIME 2026, the balanced version averages about 19,500 output tokens per problem, against 31,200 for Qwen3.5-9B, or 38 per cent fewer, while scoring 4.2 points lower. Frontier models use far fewer tokens. webAI cites about 5,000 per AIME 2026 problem for OpenAI's GPT-5.5, and it says the comparisons are directional because testing setups and tokenizers differ.

TwIL-8B extends a family that began with 1.7-billion and 3-billion-parameter formal-logic models released in August 2026. The model has a 131,072-token context window and handles English text only. No hosted inference provider serves it on Hugging Face yet, so for now users have to run it themselves.