Built for compliance rules, contract logic, and research reasoning, the 3B model beats OpenAI's open-weights gpt-oss-120b, a model 40× its size, on four of five formal-reasoning benchmarks, while the 1-gigabyte 1.7B variant outperforms every sub-2B model webAI evaluated.
[AUSTIN, Texas] — August 10, 2026 webAI today released TwiL-LM (Thinking webAI Intelligence Lab Language Model), a family of small formal-logic models at 1.7 billion and 3 billion parameters that run entirely on consumer hardware and outscore far larger models at deductive reasoning. TwiL-LM3, the 3B model, beats gpt-oss-120b, a model 40 times its size, on four of five benchmarks in webAI's formal-reasoning suite. Both models are available now on Hugging Face: TwiL-LM1.7B, TwiL-LM3B.
TwiL-LM does one hard thing exceptionally well: it translates plain English into formal logic, checks whether conclusions follow from their premises, and works through multi-step deductive reasoning which is the backbone of compliance rules, contract conditions, research, and everyday decision-making. This is a contribution to the sub-field known as autoformalization.
The AI-reasoning race has so far been a story about scale, with frontier labs pursuing better reasoning through ever-larger models served from data centers. TwiL takes the opposite bet: pure deductive reasoning, distilled into a package small enough to download onto a phone at about one gigabyte to connect with your non-internet data sources. It is an expert an enterprise can own and run at the edge instead of renting through an API.
“What’s surprised me most about TwiL is how useful it is beyond any single benchmark,” said David Stout, CEO and co-founder of webAI. “I use it every day for writing, tool calling, and general reasoning, but the more interesting capability is how well it works alongside other expert models. TwiL can act almost like an auto-correct for AI — checking outputs, refining reasoning, enforcing structure, and helping specialized models produce better answers. And because it runs at around 300 tokens per second on an M2 MacBook, you can use that reasoning layer continuously without slowing everything else down.”
The numbers
Against gpt-oss-120b on webAI’s formal-reasoning suite, TwiL-LM3 (3B) scores:
- 96.4 vs. 65.2 on rule induction (deriving rules from data)
- 87.6 vs. 43.3 on semantic parsing (language into structured queries (token F1)
- 64.6 vs. 63.1 on Lean formalization (theorem statements into formal code (token F1))
- 52.0 vs. 7.0 on exact-format answering: a 7.4× gap
- 68.7 vs. 77.5 on entailment labeling (the one lane the 120B model keeps a lead)
Beyond the formal-reasoning suite, TwiL-LM3 stays within striking distance of models many times its size on general benchmarks such as LogicBench (71.7) and GSM8K (87.3). In webAI's throughput tests it answered 2.6× faster than gpt-oss-120b, at 32.9 vs. 12.6 answers per second. The 1.7B variant, built for phones, roughly doubles its base model (0.361 vs. 0.185 aggregate reasoning accuracy over SmolLM2-1.7B) and led every sub-2B model webAI evaluated, outscoring Phi-4-mini, SmolLM-3B, Qwen2.5-3B, and Llama-3.2-3B, models in the 3-to-4-billion-parameter class.
Expertise, not scale
TwiL-LM1.7B's performance was driven by a proprietary reasoning data engine built from open sources and targeted fine-tuning from a 289 MB LoRA adapter of roughly 72 million parameters.
The release is the continued public demonstration of webAI’s collaborative intelligence thesis: that the future of AI is not one giant generalist rented through an API, but specialized AI models as experts that teams build and run themselves.
“In the quest for artificial general intelligence (AGI) I think we are relearning the importance of narrow, specialized models that capture true expertise,” said Dr. Paul J. Maykish, Chief Intelligence Officer, webAI. “A minor consensus is growing that real work gets done by teams of specific AI models that interact and improve with your non-internet, expert data. TwiL-LM1.7B is what one of those deductive-reasoning experts looks like — in a form that downloads at 1.06 GB onto your phone. We imagine that many specialized expert models combining on a network without the need for special silicon is one path to AGI.”
Built to run where the work happens
The recommended quantized build is a 1.06 GB download that ran at roughly 367 tokens per second in webAI's evaluation, fast enough for a small laptop or a phone. Nothing is sent to an external cloud, which matters for regulated environments in healthcare, financial services, and pharmaceuticals where data cannot leave the device.
Additional results include 0.655 on “entailment labeling” (checking if statements follow logically), the model's strongest objective, and 0.590 on the held-out LogicBench (a test for logic puzzles) against 0.563 for the base model. The model has an 8,192-token context window. For high-stakes formal work, webAI recommends pairing TwiL-LM with verification tooling such as a symbolic solver to benefit from the speed but account for the shorter context window. The 3B model has a longer ~65k-token context window for use cases where a standalone model is the best solution.
Availability
TwiL-LM 1.7B and 3B variants are available now on Hugging Face, in formats for both the Transformers ecosystem and local llama.cpp deployment, under the webAI Non-Commercial License v1.0.
About webAI
webAI is the enterprise AI platform that brings AI to your data. Built for mission-critical environments, it enables organizations to build and operate private, custom models with complete data sovereignty, real-time performance, and predictable economics.
Media Contact webai@pinkston.co