TwIL-8B is webAI's 8.8B reasoning model for math, code and instruction following, post-trained from IBM's Granite 4.2 8B with full-parameter SFT and then GRPO/DAPO reinforcement learning. It ships in two variants with the same architecture and tokenizer. The verbose variant is the output of the RL stage. The default balanced variant adds a length-aware RL stage for shorter reasoning traces. It thinks in a <think> block by default and needs a large generation budget (about 19.5k tokens per AIME 2026 problem), and the context window is 131,072 tokens. The card's comparison against Qwen3.5-9B, Gemma 4 12B and Gemma 4 31B shows the highest MATH-500 (98.2), AIME 2025 (88.3) and LiveCodeBench (81.6, measured on the verbose variant) of the group. It leads Qwen3.5-9B on IFBench by 9.9 points and is roughly level with it on contest math while using about 38% fewer tokens on AIME 2026. It trails on knowledge benchmarks, with GPQA Diamond at 70.2 against Qwen3.5-9B's 81.7, and well behind the frontier models (GPT-5.5, Claude Opus 4.8, Grok 4.3) on HLE and the hardest contest set, HMMT Feb 2026 (68.2 against about 96–99). Scores are from webAI's own evaluation and competitor figures are published numbers, not re-runs. Rows marked † were run only on the verbose variant, and the card says there is no safety, preference or tool-use training beyond what Granite 4.2 carries. It is available in bf16 safetensors and in GGUF builds, including a 4.98 GiB Q4_K_M, and is released under the webAI Non-Commercial License 1.0, not the Apache 2.0 license of its base model.
