TwIL
webAI's reasoning models: TwIL-8B for contest math and code, and two 3B-class formal-logic models, all non-commercial only.
In short. TwIL is a small family of open AI models from webAI that think step by step before they answer. TwIL-8B is aimed at contest-level math, code and following instructions, and two smaller models, TwIL-LM3-Pro and TwIL-LM3, are aimed at formal logic. All of them are free for non-commercial research and education only.
Summary
webAI is a company that makes a local-AI app, and TwIL is its line of open models, all post-trained from other open models. TwIL-8B first appeared on Hugging Face on 14 October 2026. It is built from IBM's Granite 4.2 8B with full-parameter supervised fine-tuning followed by GRPO/DAPO reinforcement learning, and it is an 8.79B dense model with a 131,072-token context that writes a thinking block before it answers. Two builds share one architecture, a balanced one in the repository root and a verbose one in a subfolder that reasons for longer. The two smaller models target formal logic, meaning translation into first-order logic, entailment labels, semantic parsing and Lean statements. TwIL-LM3-Pro has 3.66B parameters and a 131,072-token context and is built from Granite 4.2 3B. The earlier TwIL-LM3 has 3.08B parameters and a 65,536-token context and is built from Hugging Face's SmolLM3-3B.
For TwIL-8B, webAI reports 98.2 on MATH-500, 88.3 on AIME 2025 and 88.3 on AIME 2026 for the balanced build, and 81.6 on LiveCodeBench for the verbose build, against 65.6 for Qwen3.5-9B and 80.0 for Gemma 4 31B. It trails Qwen3.5-9B on knowledge, with 70.2 on GPQA Diamond against 81.7. About 19.5K output tokens per AIME 2026 problem is 38% fewer than Qwen3.5-9B's 31.2K, though the card puts GPT-5.5 at about 5K. For the formal-logic models, webAI reports that TwIL-LM3-Pro lifted its in-domain gate score from 0.431 to 0.554 over its Granite base while its held-out average stayed level, and that TwIL-LM3 lifted its SmolLM3 base from 0.336 to 0.422.
All scores come from webAI's own runs. For TwIL-8B the competitor figures are copied from public sources and were not re-run, so prompts, sampling and token budgets differ, and the card notes that the AIME and HMMT sets have only 30 to 33 problems, so gaps of about three points are within noise. The weights are about 17.6 GB in bf16 for TwIL-8B, 7.3 GB for TwIL-LM3-Pro and 6.2 GB for TwIL-LM3, with Q4_K_M GGUF builds of 5.35, 2.24 and 1.92 GB. In the 30 days to 11 November 2026, Hugging Face counted about 260,000 downloads for TwIL-LM3, 47,000 for TwIL-8B and 15,000 for TwIL-LM3-Pro. The license is the webAI Non-Commercial License 1.0, which is stricter than the Apache 2.0 license of the Granite and SmolLM3 base models.
Good for
- Research and teaching on contest-style math, code and instruction following at 8B, with a 5.35 GB Q4_K_M build for llama.cpp
- Formal-logic work such as first-order-logic translation, entailment labeling and Lean statements with the two smaller models
- Laptop-class machines, since the Q4_K_M files of the 3B-class models are about 2 GB
- Studying how a small reasoning model is post-trained, because the cards describe each stage, from SFT and GRPO to weight merging
Watch out for
- The license covers non-commercial research and education only, so these models cannot go into products or paid services even though their Granite and SmolLM3 bases are Apache 2.0
- TwIL-8B is not a knowledge model. It trails Qwen3.5-9B by 11.5 points on GPQA Diamond, and the card says it has no safety, preference or tool-use training beyond what Granite 4.2 carries
- It thinks at length. Even the balanced build spends about 19.5K tokens on an AIME 2026 problem, and the card warns that a short generation budget truncates the reasoning and scores far worse
- Several TwIL-8B scores, including LiveCodeBench and IFBench, were measured on the verbose build only, and the card says the balanced build has not been run on them yet
- The 3B-class models are tuned for formal logic only. Code generation and tool use were not evaluated, and every score was measured inside an 8,192-token window even though they inherit a 64K or 128K context
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the TwIL-LM3 (Q4_K_M GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| TwIL-LM3 BF16 | 3.08B | 7.08 GB | 3.54 GB | 2.13 GB | Laptop or phone |
| TwIL-LM3 (Q4_K_M GGUF) INT4 | 3.08B | 7.08 GB | 3.54 GB | 2.11 GB | Laptop or phone |
| TwIL-LM3-Pro BF16 | 3.66B | 8.42 GB | 4.21 GB | 2.53 GB | Laptop or phone |
| TwIL-LM3-Pro (Q4_K_M GGUF) INT4 | 3.66B | 8.42 GB | 4.21 GB | 2.46 GB | Laptop or phone |
| TwIL-8B (balanced) BF16 | 8.79B | 20 GB | 10 GB | 6.07 GB | Laptop or phone |
| TwIL-8B (balanced, Q4_K_M GGUF) INT4 | 8.79B | 20 GB | 10 GB | 5.88 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A short license that webAI, Inc. puts on its TwIL models. It grants a perpetual, worldwide, royalty-free copyright license to use, modify and share the weights, but only for non-commercial research and educational purposes. The same limit follows every derivative work.
- Attribution or model name must be shown
More about webAI Non-Commercial
Frequently asked questions
- What is TwIL?
- TwIL is a small family of open AI models from webAI that think step by step before they answer. TwIL-8B is aimed at contest-level math, code and following instructions, and two smaller models, TwIL-LM3-Pro and TwIL-LM3, are aimed at formal logic. All of them are free for non-commercial research and education only.
- Can I use TwIL commercially?
- Not without separate permission. TwIL is released under the webAI Non-Commercial, which allows research and personal use only.
- Is TwIL open source?
- TwIL is open-weight, not open source in the strict sense. You can download the model files, but the webAI Non-Commercial adds conditions that the Open Source Initiative definition does not allow.
- Can I run TwIL on my own computer?
- Yes. The smallest variant, TwIL-LM3 (Q4_K_M GGUF), needs about 2.11 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was TwIL released?
- TwIL was released by webAI on October 14, 2026.
- How long a context window does TwIL have?
- TwIL supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- TwIL-8B model card (Hugging Face) huggingface.co
- TwIL-LM3-Pro model card (Hugging Face) huggingface.co
- TwIL-LM3 model card (Hugging Face) huggingface.co
- webAI Non-Commercial License ver. 1.0 (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Granite 4.2 IBM | 3.7B – 29.3B | 128K | Apache 2.0 (Permissive) | 2.55 GB Laptop or phone | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Nanbeige4.2-3B Nanbeige LLM Lab | 4.17B | 256K | Apache 2.0 (Permissive) | 2.88 GB Laptop or phone |