WebAI, a company based in Austin, released its TwiL-LM family of formal-logic reasoning models on August 10. The flagship TwiL-LM3, a 3 billion parameter model, outperformed OpenAI’s gpt-oss-120B on four out of five benchmarks, including a score of 96.4 accuracy on rule induction tasks versus 65.2 for OpenAI’s model. TwiL-LM3 processed roughly 32.9 answers per second, about 2.6 times faster than OpenAI’s model, while running on consumer hardware. The TwiL-LM family comes in 1.7 billion and 3 billion parameter sizes on Hugging Face in GGUF format under a non-commercial license. WebAI targets deductive reasoning applications such as compliance checks and contract analysis.
Source: Read the original article
