webAI puts a 3B logic model on consumer hardware

webAI released TwiL-LM3, a 3-billion-parameter formal-logic model that beat the 120-billion-parameter gpt-oss model on four of five company benchmarks. Its 1.7B sibling downloads at 1.06 GB and reportedly runs at 367 tokens per second, but both models carry a noncommercial license and the benchmark suite is vendor-run. The practical lesson echoes Kimi K3’s open-weight cost case: narrow local experts can compress inference, yet licensing and independent evaluation must clear before production routing does.