README.md

---
title: TwIL-LM3-Pro
emoji: 🧠
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
pinned: true
license: other
short_description: Official demo of the TwIL-LM3-Pro logic reasoner
python_version: "3.12"
startup_duration_timeout: 30m
---

# TwIL-LM3-Pro — Official Demo

Official Hugging Face Space for
[`webAI-Official/TwIL-LM3-Pro`](https://huggingface.co/webAI-Official/TwIL-LM3-Pro),
webAI's 3.66B reasoning model for **formal logic**.

TwIL-LM3-Pro is built from
[`ibm-granite/granite-4.2-3b`](https://huggingface.co/ibm-granite/granite-4.2-3b)
through LoRA SFT, checkpoint fusion, WiSE-FT interpolation, and entropy-weighted
GRPO. It emits a collapsible `<think>…</think>` trace before the answer.

**Temperature = 0** (greedy) matches the published evaluation protocol. Keep
max new tokens at 2048 or higher so the reasoning block is not truncated. The
interactive control supports up to **16,384 new tokens** for long reasoning
traces; larger requests reserve more ZeroGPU time and consume more visitor
quota.

This Space runs on ZeroGPU and streams tokens as they are generated.

> **License:** the model is released under the *webAI Non-Commercial License
> ver. 1.0* — see the [model repository](https://huggingface.co/webAI-Official/TwIL-LM3-Pro)
> for full terms.