webAI releases TwIL-8B, an 8.8B reasoning model that runs on a laptop

TwIL-8B went through full-parameter supervised fine-tuning, then GRPO/DAPO reinforcement learning, then a final length-aware RL stage that trims how long it thinks[1]. That produces two variants sharing one architecture and tokenizer: "balanced", the default, and "verbose", which thinks longer[1]. It's a 40-layer dense decoder with a 131,072-token context window inherited from Granite, English text in and out[1].

In webAI's comparison against open models of similar size and up to 3.5 times larger, TwIL-8B has the top MATH-500 score at 98.2, against 96.6 for Qwen3.5-9B, and the top LiveCodeBench score at 81.6, against 80.0 for Gemma 4 31B[1]. It scores 74.4 on IFBench, 9.9 points above Qwen3.5-9B[1]. The LiveCodeBench and IFBench figures were measured on the longer-thinking verbose variant, not the default balanced one[1]. The balanced variant also uses about 19.5k output tokens per AIME 2026 problem, 38% fewer than Qwen3.5-9B's 31.2k[1].

The weak spot is knowledge. TwIL-8B gets 70.2 on GPQA Diamond, 11.5 points behind Qwen3.5-9B, and 12.6 on Humanity's Last Exam with the verbose variant[2]. The competitor numbers were taken from public sources rather than re-run, so treat the comparison as webAI's own[1].

The full bf16 checkpoint is 16.38 GiB, and a 4.98 GiB Q4_K_M GGUF build is sized for CPUs or modest-VRAM GPUs through llama.cpp[2]. The model needs a big generation budget: the card's example sets 32,768 new tokens, and short budgets cut its reasoning off and hurt scores[2]. The licence is the webAI Non-Commercial License 1.0, so it's for research and personal use rather than products[1].

webAI already ships smaller TwIL models in its local desktop app, where TwIL 3B Pro is the default on 8 GB Macs[3].

Small reasoning models often get the right answer but spend tens of thousands of tokens getting there, which is slow on local hardware. TwIL-8B's length-aware training is a direct attack on that, and a 5 GB model that scores near frontier systems on maths and code makes local reasoning more practical, as long as you don't need it commercially or for knowledge-heavy questions.