webAI Releases Local Formal-Logic Models in 1.7B and 3B Sizes
webAI has released TwIL-LM, a two-model family designed to translate English into first-order logic and determine whether conclusions follow from stated premises. The models contain 1.7 billion and 3 billion parameters and are intended to run locally. The 1.7B quantized version is 1.06 GB, while the 3B Q4_K_M GGUF build is 1.78 GiB and can run on a CPU or 4 GB of VRAM.
The 3B model, TwIL-LM3, is based on a merged fine-tune of SmolLM3-3B. The smaller model uses a PEFT LoRA adapter for SmolLM2-1.7B-Instruct. Potential uses include structured-query generation, Lean formalization drafting, entailment classification and checking the output of larger models. webAI reported scores of 96.4 for rule induction, 87.6 for semantic parsing and 64.6 for Lean formalization.
On the reported six-lane average, TwIL-LM3 scored 0.4488, below gpt-oss-120b’s 0.5192. Its generation length was 482 tokens, with 32.9 answers per second versus 4.2 for the 120B model. Both checkpoints are currently covered by the webAI Non-Commercial License ver. 1.0; revenue-generating deployment requires a separate agreement.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.