Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
Base Model, Architecture, and the Two Variants
TwIL-8B is an 8.79B-parameter dense decoder-only model post-trained from IBM's Granite 4.2 8B. It has 40 layers, a hidden size of 4096, and 32 attention heads with 8 KV heads, and it handles English text input and output only. Its 131,072-token context is inherited from the base model. Weights are on Hugging Face under the webAI Non-Commercial License 1.0; the base model is Apache 2.0.
webAI's earlier TwIL models were 1.7B and 3B formal-logic models; TwIL-8B moves the target to contest math, code, and instruction following. Before answering, the model writes a `<think>` reasoning block. The two variants share the same architecture and tokenizer: balanced is the default with shorter traces, while verbose reasons longer and targets harder contest-style problems. The 4-bit Q4_K_M GGUF is 4.98 GiB, small enough for CPU or modest-VRAM GPU inference; bf16 weights are 16.38 GiB per variant. The model card also warns that a short generation budget truncates reasoning; its example command uses a 40960 context and a 32768 generation cap.
How Three Post-Training Stages Produce the Two Versions
Training runs in three stages. First, full-parameter supervised fine-tuning on reasoning traces for math, STEM, and instruction following. Second, GRPO/DAPO reinforcement learning on math, code, STEM, and instruction prompts, which produces the verbose variant. Third, length-aware reinforcement learning, where the reward favors shorter traces against per-domain length references, producing the default balanced variant.
The model card says training pools were decontaminated with 13-gram matching against evaluations including AIME, HMMT, MATH-500, and GPQA. Public material does not disclose the full data list, per-stage data volumes, training compute, or the exact reward design of the length-aware RL stage.
Benchmarks: Strong on Math and Code, Behind on Knowledge
All TwIL-8B scores are self-reported by webAI as high-effort pass@1 with thinking enabled; scores marked † were measured on the verbose variant only. On math, MATH-500 is 98.2, against Qwen3.5-9B's 96.6 on a third-party leaderboard and Gemma 4 31B's 87.9 in a Red Hat lighteval run. AIME 2026 is 88.3, below Qwen3.5-9B's 92.5 and Gemma 4 31B's 89.2; AIME 2025 is 88.3, slightly above Gemma 4 31B's 87.1.
On code, LiveCodeBench† is 81.6 versus Qwen3.5-9B's 65.6 and Gemma 4 31B's 80.0. But the model card notes that the version of LiveCodeBench webAI used was not recorded, while competitor scores are v6, so this comparison is loose. On instruction following, IFBench† is 74.4, 9.9 points above Qwen3.5-9B's 64.5; IFEval is essentially level at 91.8 versus 91.5.
Knowledge is the clear weak spot. GPQA Diamond is 70.2, 11.5 points behind Qwen3.5-9B's 81.7 and also below Gemma 4 31B's 84.3; MMLU-Pro† is 76.5 against Qwen3.5-9B's 82.5. Across 13 shared benchmarks, TwIL-8B goes 7-6 against Qwen3.5-9B and 3-11 against Gemma 4 31B, which is about 3.5 times its size. Competitor figures are each vendor's published numbers; webAI did not re-run them.
How Many Tokens It Saves, and What That Costs
By the model card's mean output tokens per problem, TwIL-8B uses about 19.5k on AIME 2026 versus 31.2k for Qwen3.5-9B, 38% fewer, at 4.2 points lower accuracy. On HMMT Feb 2026 the gap is 32.9k versus 40.6k, 19% fewer, at 3.0 points lower accuracy. webAI also cites about 5k tokens per AIME 2026 problem for GPT-5.5 as a reference point, but states that harnesses and tokenizers differ, so that comparison is directional only.
What these numbers show is that length-aware RL trims reasoning cost within webAI's own model at a cost of a few accuracy points. They do not support the claim that TwIL-8B is cheaper to run than frontier models.
Three Things to Check Before Adopting It
First, task type. TwIL-8B is built for verifiable computational work such as math and code; knowledge Q&A is not its ground, as the GPQA Diamond and MMLU-Pro gaps show. The model card mentions no additional safety, preference, or tool-use training, so treating it as a general assistant or relying on it for tool calls is a mismatch.
Second, the generation budget. The model card warns that a short budget truncates reasoning and drops scores, so deployment should set a token cap high enough for the hardest expected problem. Use balanced by default to control cost and switch to verbose only for hard contest-style inputs.
Third, the license. Public material does not spell out the specific terms of the webAI Non-Commercial License 1.0, so teams considering commercial integration need to read the license text directly.