TwIL-8B

TwIL-8B is an 8.8B reasoning model open-sourced on 2026-11-04 by Austin-based webAI (CEO David Stout). It starts from ibm-granite/granite-4.2-8b, then full-parameter SFT + GRPO/DAPO RL, plus length-aware RL that yields balanced (default) and verbose variants. The card claims top MATH-500 (98.2) and LiveCodeBench (81.6, verbose) among similar open peers, IFBench ~9.9 above Qwen3.5-9B, and roughly 88–93% of frontier on coding/instruction rows.

Weights and GGUFs live at Hugging Face webAI-Official/TwIL-8B (~16.4 GiB bf16; ~5 GiB Q4_K_M). License: webAI Non-Commercial License v1.0 (Granite base remains Apache 2.0). Note: the existing WebAI lead in this DB is Opera’s browser WebAI API—not the same company as webAI.com’s TwIL line.

Problem

Small local models often lag on contest math and coding. TwIL-8B targets those axes and uses length-aware RL to shorten traces.

Differentiation

Same TwIL family as TwIL-LM3 / TwIL-LM3-Pro, but the 8B line is math/code/IF—not pure formal logic. Versus Qwen3.5-9B: stronger code/IF, weaker knowledge (GPQA etc.).

Limits

• Non-commercial license for the release weights.
• Knowledge/HLE lag larger models; traces remain long.
• Some † scores are verbose-only so far.

Who it’s for

Researchers and individuals wanting local contest-math/code reasoning under a non-commercial license.

Get started

1. Open the HF card.
2. Load via Transformers or llama-cli with the balanced Q4 GGUF.
3. Use large generation budgets; pick balanced vs verbose per task hardness.