webAI Introduces TwIL-LM: A Formal-Logic Model Family for Local Autoformalization

webAI released TwIL-LM, a two-model family of formal-logic reasoners at 1.7B and 3B parameters. The 3B member, TwIL-LM3, is a merged fine-tune of SmolLM3-3B; the 1.7B member is a PEFT LoRA adapter for SmolLM2-1.7B-Instruct. Both target autoformalization: translating English into first-order logic and checking whether a conclusion follows from its premises. Both run locally, with a 1.06 GB quantized build for the 1.7B and a 1.78 GiB Q4_K_M GGUF for the 3B. webAI’s announcement frames the release around beating gpt-oss-120b on four of five formal-reasoning lanes. The deployment is partial, as both checkpoints ship under the webAI Non-Commercial License ver. 1.0; revenue-generating deployment requires a separate agreement. The models target various industries including compliance and RegTech, financial services, healthcare and pharma, legal and contract operations, and formal-methods research. Applications include first-order logic (FOL) translation, entailment classification, natural language to structured query, Lean formalization drafting and critique, and a verifier layer. TwIL-LM3 was built by four stages: LoRA supervised fine-tuning, checkpoint fusion, WiSE-FT interpolation back toward the pretrained base at λ = 0.25, and MGPO, an entropy-weighted GRPO stage run against a programmatic verifier. The published checkpoint is step 2071. TwIL-LM3 scores 96.4 on rule induction, 87.6 on semantic parsing, 64.6 on Lean formalization, 52.0 on exact-format answering, and 68.7 on entailment labeling. On Track A, it scores 0.4488 on the six-lane average and 0.4218 on the macro gate, leading every arm up to LFM2.5-8B-A1B but not the two largest arms (Qwen3-8B and gpt-oss-120b). Efficiency is a key edge: TwIL-LM3 produces the shortest generations (482 tokens) and most answers per second (32.9 vs 4.2 for the 120B). TwIL-LM3 improves in-domain by +26% relative while also gaining +0.022 on held-out core average, being the only arm to gain on both tracks. The 1.7B model offers a different trade with mixed out-of-distribution results. Key takeaways: TwIL-LM3 and TwIL-LM are for formal logic under non-commercial license; TwIL-LM3 trails gpt-oss-120b on the six-lane average; its real edge is efficiency; WiSE-FT at λ=0.25 is why in-domain gains do not collapse held-out performance.