A model with 3 billion parameters just embarrassed one with 120 billion. WebAI, an Austin-based AI company, released its TwiL-LM family of formal-logic reasoning models on August 10, and the benchmarks tell a story that should make anyone paying cloud compute bills a little uncomfortable.
The flagship TwiL-LM3, a 3B-parameter model, outperformed OpenAI’s gpt-oss-120B on four out of five benchmarks in webAI’s formal-reasoning suite. That’s a model roughly 40 times smaller beating a giant on nearly every test thrown at it. On rule induction tasks specifically, TwiL-LM3 scored 96.4 accuracy compared to 65.2 for the OpenAI model.
Small model, big throughput
Raw accuracy is only half the picture. TwiL-LM3 also processed roughly 32.9 answers per second, compared to approximately 12.6 for OpenAI’s model. That’s about 2.6 times faster throughput from a model that can actually run on hardware you already own.
The TwiL-LM family ships in two sizes. The 1.7B-parameter variant is optimized for phones, targeting on-device inference for mobile applications. The larger 3B TwiL-LM3 is the benchmark-topping version, fine-tuned using SmolLM3-3B as a base model and enhanced with LoRA adapters. LoRA, short for Low-Rank Adaptation, is a technique that lets developers fine-tune large models efficiently by only training a small number of additional parameters rather than retraining the entire network.
