webAI open-sources TwIL-8B, an 8.8B model built on Granite
An 8.8B model matching near-frontier coding and math quality at consumer-hardware size changes the calculus for local inference, where running Claude- or GPT-class ability previously required cloud APIs. Developers building offline or cost-sensitive tools can deploy it from a single quantized 4.98 GiB file on commodity machines. Its efficiency-trained reasoning also cuts token counts, which matters for anyone benchmarking small models against Qwen3.5-9B or Gemma 4 31B.
webAI open-sourced TwIL-8B, an 8.8B model fine-tuned from IBM's Granite 4.2 8B for math, coding and instruction following. The quantized model fits in a 4.98 GiB file and runs on a CPU or modest GPU. It scores 81.6% on LiveCodeBench, 16.0 points above Qwen3.5-9B and ahead of Google's Gemma 4 31B. On IFBench it beats Claude Opus 4.8 by 12.2 points, reaching 93% of the best frontier score on coding against GPT-5.5, Claude Opus 4.8 and Grok 4.3. A training stage rewarding shorter reasoning yields a 98.2% MATH-500 score and 38% fewer tokens than Qwen3.5-9B on AIME 2026.
webAI (@thewebAI) just open-sourced TwIL-8B, an 8.8B model built on IBM’s Granite 4.2 8B and trained for math, coding and following instructions.
> TwIL-8B fits in a 4.98 GiB quantized file that runs on a CPU or modest graphics card, putting near-frontier coding and math on hardware people already own.
> scores 81.6% on LiveCodeBench, 16.0 points above Qwen3.5-9B and ahead of Google’s Gemma 4 31B, a model 3.5x its size.
> Against GPT-5.5, Claude Opus 4.8 and Grok 4.3, TwIL-8B reaches 93% of the best frontier score on coding and 92% on instruction following, beating Claude Opus 4.8 on IFBench by 12.2 points.
> A final training stage that rewards shorter reasoning lets TwIL-8B solve AIME 2026 problems with 38% fewer tokens than Qwen3.5-9B, while posting the highest MATH-500 score in its comparison at 98.2%.
