webAI opens a local 8.8B reasoning model — but bans commercial use
TwIL-8B targets coding, math and instruction following on modest hardware, while its custom license limits use to non-commercial research and education.
webAI has released TwIL-8B, an 8.79-billion-parameter reasoning model tuned for math, coding and instruction following. Its weights and quantized GGUF builds are available on Hugging Face, including a 4.98 GB version intended for CPUs and GPUs with modest memory.
The model is based on IBM’s Granite 4.2 8B and supports a 131,072-token context window. webAI trained it with supervised fine-tuning and reinforcement learning, then published two variants: a balanced default that aims to reduce output length and a verbose version for harder contest-style problems.
webAI reports a score of 81.6 on LiveCodeBench for the verbose variant, ahead of the published figures it cites for Qwen3.5-9B and Gemma 4 31B. It also reports 98.2 on MATH-500 and strong instruction-following results. The comparisons are not neutral evaluations: competitor numbers were not re-run, several TwIL scores use the verbose variant, and the model card says TwIL remains weaker on knowledge-heavy tests such as GPQA Diamond.
The biggest caveat is licensing. Although webAI calls the release open source, the custom license permits use only for non-commercial research and education. It also requires derivatives to keep the same non-commercial restriction. The base Granite model uses Apache 2.0, but those broader rights do not carry over to TwIL-8B.
That makes TwIL useful for local experimentation, evaluation and education, but not a drop-in model for commercial products. The release still shows how far compact models can go on coding and math without requiring datacenter-class hardware.
