webAI Releases TwIL-LM: A 1.7B and 3B Formal-Logic Model Family for Autoformalization on Local Hardware
webAI drops TwIL-LM: 1.7B and 3B models that translate English to formal logic, run on CPU. But the benchmark scores? From an unreleased checkpoint.
Why it matters
A specialized open-weight model family for autoformalization (formal-logic reasoning) that runs locally on modest hardware. Relevant to practitioners building reasoning pipelines and those tracking the fragmentation of narrow-task models. The caveat: benchmark claims don't match released weights.
The key facts
8 to knowTwo model sizes: 1.7B (1.06GB download) and 3B (runs on 4GB VRAM or CPU)
Task: English-to-first-order-logic translation and logical entailment checking
Non-commercial license
Headline benchmark scores attributed to unreleased checkpoint, not Hub weights
Specialist model for autoformalization and formal reasoning
Task: English-to-first-order-logic translation and logical inference (premise-to-conclusion checking)
Critical caveat: headline benchmark scores from unreleased checkpoint, not the published weights
Domain: autoformalization and formal reasoning
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: webAI has released TwIL-LM, a family of formal-logic models at 1.7B and 3B parameters that translate English into first-order logic and check whether conclusions follow from premises. The 3B runs on CPU or 4GB of VRAM; the 1.7B downloads at 1.06GB. Both ship under a non-commercial license. The…