Editorial illustration for webAI releases TwIL-8B, an 8.8B reasoning model scoring 98.2 on MATH-500
webAI Releases TwIL-8B Reasoning Model, Scores 98.2
webAI put out TwIL-8B on Hugging Face this week, an 8.8-billion-parameter reasoning model built on top of IBM's Granite 4.2 8B. The company's own benchmark table has it hitting 98.2 on MATH-500, the top score among the models it compared against, and the release comes with the usual mix of strong and weak spots that define small reasoning models right now. On GPQA Diamond it lands at 70.2, trailing Qwen3.5-9B by 11.5 points, and its Humanity's Last Exam score sits at 12.6 for the verbose variant.
The model ships in two flavors, Balanced and Verbose, both sharing the same 40-layer architecture, 4096 hidden size, and 131,072-token context window inherited from Granite. A 4-bit GGUF build weighs in at 4.98 GiB, small enough to run on a laptop CPU or a GPU with modest VRAM, while the full bf16 weights run 16.38 GiB. Weights are available under a non-commercial license.
TwIL-8B is the third release in webAI's TwIL line, following a pair of smaller formal-logic models, 1.7B and 3B, that came out in August. The company's pitch this time centers less on raw scale and more on what each task costs to run.
webAI has released TwIL-8B, an 8.8B reasoning model for math, code and instruction following. The TwIL-8B reasoning model is post-trained from IBM’s Granite 4.2 8B. Weights are on Hugging Face under a non-commercial license.
Why this matters
A score of 98.2 on MATH-500 from an 8.8B model sounds absurd until you remember these numbers are self-reported, pass@1 with thinking enabled, and compared against a Gemma 4 31B run that Red Hat put at 87.9. webAI's own framing, cost per task rather than raw capability, tells you where they want the conversation to go: a 4.98 GiB quantized build that runs on CPU or a modest GPU is a real deployment story for teams tired of renting A100s for every inference call. The 13-gram decontamination against AIME, HMMT, GPQA and others is a sensible disclosure, but it's also the kind of thing you want a third party to verify, not just take on faith from a launch post.
Granite 4.2 as the base matters too: IBM's lineage plus a non-commercial license on Hugging Face means this is a research artifact first, a product second. For builders evaluating small reasoning models, TwIL-8B is worth benchmarking yourself before betting a roadmap on numbers nobody outside webAI has reproduced yet.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Brennan: OpenAI's 2028 Election Policy Assumes Mass AI Adoption Not Yet Seen
From Archives Hugging Face CEO Says P(Doom) Close to Zero; 20M AI Builders, 50K Microducks
Common Questions Answered
What is TwIL-8B and what base model was it built from?
TwIL-8B is an 8.8-billion-parameter reasoning model released by webAI that is specifically designed for math, code, and instruction following tasks. The model was post-trained from IBM's Granite 4.2 8B and is available on Hugging Face under a non-commercial license.
How does TwIL-8B's MATH-500 score compare to other models?
TwIL-8B achieved a score of 98.2 on MATH-500, which is the top score among the models webAI compared in their benchmark table. This score outperforms Gemma 4 31B, which scored 87.7 on the same benchmark, making the smaller model's performance particularly noteworthy.
What are TwIL-8B's performance metrics on other benchmarks besides MATH-500?
On GPQA Diamond, TwIL-8B scores 70.2, which trails Qwen3.5-9B by 11.5 points. Additionally, its Humanity's Last Exam score sits at 12.6 for the verbose variant, demonstrating that while strong in mathematical reasoning, the model shows weaker performance on other challenging benchmarks.
What are the practical deployment advantages of TwIL-8B's size?
TwIL-8B's 4.98 GiB quantized build can run on CPU or modest GPU hardware, making it a practical deployment option for teams looking to avoid expensive A100 GPU rental costs. This efficiency makes the model particularly valuable for organizations focused on cost per task rather than raw computational capability.
Why is TwIL-8B's MATH-500 score of 98.2 considered remarkable?
The score is notable because it comes from an 8.8B parameter model using pass@1 scoring with thinking enabled, which are self-reported metrics from webAI's own benchmark. The fact that this small model outperforms much larger models like Gemma 4 31B demonstrates significant efficiency gains in reasoning capability relative to model size.
Further Reading
- TwIL-8B: 8.8B Model Claims 93% of Frontier Coding Score - Explainx.ai
- webAI-Official/TwIL-8B - Hugging Face
- Granite 4.2 - IBM Granite
