- Published on
A 1.7B and 3B Formal-Logic Model Family for Autoformalization on Local Hardware
- Authors
- Name
- DeepHorus
webAI has unveiled TwIL-LM, a compact family of formal-logic reasoning models available in 1.7B and 3B parameter sizes. The flagship TwIL-LM3 is a merged fine-tune of SmolLM3-3B, while the 1.7B variant uses a LoRA adapter on SmolLM2-1.7B-Instruct. Both models specialize in autoformalization, translating natural English into first-order logic and verifying whether conclusions logically follow from premises. They are designed to run entirely on local hardware, with the quantized builds occupying just 1.06 GB and 1.78 GiB respectively.
This release targets environments where data cannot leave the device, making local execution a key selling point. Industries such as compliance, RegTech, financial services, healthcare, legal operations, and formal-methods research stand to benefit. The models also enable first-order logic translation, entailment classification, natural language to structured query generation, and Lean formalization drafting, including a verifier layer that checks the outputs of larger models.
webAI frames the release around competitive performance, claiming TwIL-LM3 beats gpt-oss-120b on four of five formal-reasoning lanes. On concrete benchmarks, 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. However, on the six-lane average, it trails gpt-oss-120b with 0.4488 versus 0.5192, and Qwen3-8B leads on the macro gate.
Efficiency is where the model stands out: TwIL-LM3 produces the shortest generations of any arm at 482 tokens and delivers 32.9 answers per second, compared to 4.2 from the 120B model. Its training involved LoRA supervised fine-tuning, checkpoint fusion, and WiSE-FT interpolation retaining only a quarter of the fine-tuned delta, which helped preserve held-out capability. The 1.7B variant shows more mixed out-of-distribution results, with GSM8K falling to 0.380 from 0.413.
Both checkpoints ship under the webAI Non-Commercial License ver. 1.0, meaning revenue-generating deployment requires a separate agreement. This licensing choice could limit broader adoption but keeps the models accessible for research and experimentation. The company positions these tools for formal-methods research and verification-heavy applications where accuracy and privacy trump scale.
Looking ahead, webAI's focus on formal logic at small parameter counts signals a growing trend toward specialized, efficient reasoning models. The availability of quantized builds on CPU or 4GB VRAM lowers entry barriers for individual developers and small teams. As the field matures, further iterations could close the gap with larger models while preserving local execution. For now, TwIL-LM offers a practical option for privacy-sensitive formal reasoning tasks.
