webAI Releases TwIL-LM: A 1.7B and 3B Formal-Logic Model Family for Autoformalization on Local Hardware
| Source: MarkTechPost
Tags: TwIL-LM, webAI, formal logic, autoformalization, Lean, local AI, SmolLM3, logic reasoning
webAI releases TwIL-LM, a 1.7B and 3B formal logic model family for autoformalization and entailment checking that runs under 2GB on local hardware — but ships under a non-commercial license and is outperformed by Qwen3-8B on key benchmarks despite using a third of the parameters.
Details
TwIL-LM targets a narrow but genuinely useful niche: translating natural language into first-order logic, checking whether conclusions follow from premises, and drafting Lean formalizations — all on local hardware. The 3B model (TwIL-LM3) is built from SmolLM3-3B using a four-stage training process: LoRA supervised fine-tuning on a synthetic formal-logic corpus, checkpoint fusion, WiSE-FT interpolation at λ=0.25 back toward the pretrained base, and MGPO (entropy-weighted GRPO) against a programmatic verifier.\n\nOn webAI's Track A benchmarks, TwIL-LM3 scores 0.4218 on the macro gate, ahead of models up to LFM2.5-8B-A1B. However, Qwen3-8B scores 0.5336 on the same metric with roughly 3x the parameters. The gpt-oss-120b six-lane average of 0.5192 also beats TwIL-LM3's 0.4488. The efficiency story is cleaner: at 1.78 GiB Q4_K_M, the model runs on 4GB VRAM or CPU-only hardware.\n\nCommercial deployment requires a separate license agreement with webAI. Target industries include compliance, RegTech, legal, and financial services — contexts where formal logic verification has natural value as a lightweight verifier layer for larger models.