webAI's 3.66B TwIL-LM3-Pro Beats VibeThinker by 35% in Formal Logic

Austin-based webAI released its latest reasoning model for local hardware, introducing an open-source tool designed for checking deductions and critiquing proofs. The TwIL-LM3-Pro model features 3.66 billion parameters and leads other small-scale reasoners in formal logic, scoring 35% higher than China's VibeThinker-3B, 24% higher than Qwen3.5-4B, and 47% higher than Liquid AI's LFM2.5-8B-A1B. On the GSM8K evaluation for math word problems, the model scored 94.3%, nearly matching the 97.7% result of the gpt-oss-120b which has 33 times more parameters.

The model is distributed as a 2.09GiB Q4 build that runs via llama.cpp on CPUs or local GPUs to keep private data on the device. This release follows 500,000 downloads of the company's first-generation models within a single month. webAI is now developing Meridian, a series of frontier-class on-device models to be available through the webAI application.