webAI has released TwIL-8B, an 8.79B dense decoder-only reasoning model post-trained from IBM Granite 4.2 8B, featuring 40 layers, 4096 hidden size, and 32 attention heads with 8 KV heads. The model supports a 131,072-token context and generates a reasoning trace before answering, shipping in Balanced and Verbose variants. Training involved three stages: full-parameter SFT on math, STEM, and instruction traces; GRPO/DAPO reinforcement learning producing the Verbose variant; and length-aware RL favoring shorter traces for the Balanced variant. Data was decontaminated via 13-gram matching against benchmarks including AIME, MATH-500, and GPQA. The 4-bit Q4_K_M GGUF build is 4.98 GiB for CPU or modest-VRAM GPU inference, while bf16 weights are 16.38 GiB. Self-reported high-effort pass@1 scores show TwIL-8B leading its size class on math and code: 98.2 on MATH-500, 88.3 on AIME 2025 and 2026, and 81.6 on LiveCodeBench (Verbose). It also scores 74.4 on IFBench and 91.8 on IFEval. However, knowledge benchmarks are a weak spot: GPQA Diamond is 70.2 (11.5 points behind Qwen3.5-9B's 81.7) and MMLU-Pro is 76.5. On AIME 2026, TwIL-8B averages ~19.5k output tokens versus Qwen3.5-9B's 31.2k, using 38% fewer tokens while scoring 4.2 points lower. Weights are available on Hugging Face under a non-commercial license.