webAI has released TwIL-8B, an 8.8B parameter model designed for mathematical reasoning, code generation, and instruction following. The weights are available on Hugging Face under a non-commercial license. The launch post emphasises cost per task rather than raw model scale. The 4-bit quantised build is 4.98 GiB, making it suitable for CPU inference or GPUs with limited video memory.
TwIL-8B specifications
The model contains 8.79 billion parameters in a dense architecture. It supports a 131,072-token context window inherited from its base model. TwIL-8B is a decoder-only model with 40 layers, a hidden size of 4096, and 32 attention heads, 8 of which are key-value heads. Input and output are restricted to English text. The model generates a <think> block before providing an answer.
Two variants ship with the same architecture and tokenizer. The Balanced variant is the default and produces shorter reasoning traces. The Verbose variant generates longer traces to target difficult contest-style problems.
TwIL-8B extends the TwIL family. The first members were 1.7B and 3B formal-logic models released in August 2026.
Training methodology
webAI used three post-training stages:
- Full-parameter supervised fine-tuning on reasoning traces for mathematics, STEM, and instruction following.
- GRPO/DAPO reinforcement learning on math, code, STEM, and instruction prompts to produce the Verbose variant.
- Length-aware reinforcement learning with a reward function that favours shorter traces against per-domain length references to produce the Balanced variant.
Training pools were decontaminated using 13-gram matching against AIME, HMMT, MATH-500, GPQA and other evaluation sets.
Benchmark performance
All TwIL-8B scores are self-reported, high-effort pass@1 metrics with thinking enabled. Scores marked † were measured on the Verbose variant only.
In mathematics, TwIL-8B scores 98.2 on MATH-500. Qwen3.5-9B lists 96.6 on a third-party leaderboard. Gemma 4 31B scored 87.9 in Red Hat’s lighteval run. On AIME 2025, TwIL-8B posts 88.3 against 87.1 for Gemma 4 31B. On AIME 2026 it scores 88.3, behind Qwen3.5-9B at 92.5 and Gemma 4 31B at 89.2.
For code, LiveCodeBench† is 81.6, against 65.6 for Qwen3.5-9B and 80.0 for Gemma 4 31B. webAI notes its LiveCodeBench version was not recorded. Competitor scores are v6.
Instruction following on IFBench† is 74.4, 9.9 points above Qwen3.5-9B at 64.5. IFEval is level, at 91.8 versus 91.5.
Knowledge remains the weak spot. GPQA Diamond is 70.2, against 81.7 for Qwen3.5-9B and 84.3 for Gemma 4 31B. MMLU-Pro† is 76.5 versus 82.5 for Qwen3.5-9B.
Across 13 shared benchmarks, TwIL-8B goes 7-6 against Qwen3.5-9B. Against Gemma 4 31B, about 3.5x its size, it goes 3-11.
Reasoning token usage
TwIL-8B uses fewer reasoning tokens than its closest open peer. On AIME 2026, TwIL-8B averages about 19.5k output tokens per problem. Qwen3.5-9B uses 31.2k, so TwIL-8B uses 38% fewer. It scores 4.2 points lower. On HMMT Feb 2026 the gap is 32.9k versus 40.6k, or 19% fewer.
Frontier models remain far leaner. webAI cites about 5k tokens per AIME 2026 problem for GPT-5.5. webAI states that harnesses and tokenizers differ, so treat this as directional.
Comparison table
TwIL-8B figures from the model card. † = verbose variant. All scores self-reported by each vendor unless marked.
| Metrics | TwIL-8B | Qwen3.5-9B | Gemma 4 12B | Gemma 4 31B |
|---|---|---|---|---|
| Parameters | 8.79B dense | 9B (language model) | 11.95B dense | 30.7B dense |
| Context | 131,072 | 262,144 native | 256K | 256K |
| Modality (input) | Text | Text, image, video | Text, image, audio | Text, image |
| License | webAI Non-Commercial 1.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Smallest official local build | 4.98 GiB (Q4_K_M) | Not disclosed | Not disclosed | Not disclosed |
| MATH-500 | 98.2 | 96.6 (3rd party) | Not disclosed | 87.9 (Red Hat) |
| AIME 2026 | 88.3 | 92.5 (MathArena) | 77.5 | 89.2 |
| LiveCodeBench | 81.6 † | 65.6 (v6) | 72.0 (v6) | 80.0 (v6) |
| GPQA Diamond | 70.2 | 81.7 | 78.8 | 84.3 |
| MMLU-Pro | 76.5 † | 82.5 | 77.2 | 85.2 |
Local deployment
from_pretrained(“webAI-Official/TwIL-8B”) loads the Balanced variant. Pass subfolder=”verbose” for the Verbose one. The model card lists GGUF builds for llama.cpp:
- Q4_K_M, 4.98 GiB, for both variants
- Q8_0, 8.70 GiB, Verbose only
- F16, 16.38 GiB, Verbose only
Give it a large generation budget. The card’s example sets max_new_tokens=32768. A short budget truncates reasoning and scores far worse. All reported scores used bf16 weights, so expect small drops at Q4_K_M. No hosted inference provider serves the model on Hugging Face yet.
What it means
The model targets specific tasks where efficiency matters. For users running local instances, the 4.98 GiB footprint allows operation on modest hardware. The shorter reasoning traces in the Balanced variant reduce output costs compared to peers. However, the non-commercial license limits commercial deployment. Users requiring commercial use must look elsewhere or await a license change. The Verbose variant offers better performance on hard problems but requires more memory and tokens.
