webAI releases a 3.66B formal‑logic model that runs in 2 GiB
webAI says its TwIL-LM3-Pro, a 3.66-billion-parameter model that checks whether conclusions follow from premises, roughly matches Qwen3-8B on formal logic and fits in a 2.09 GiB file.
webAI has published TwIL-LM3-Pro, a 3.66-billion-parameter model built to check whether a conclusion follows from its premises, according to the TwIL-LM3 page on Hugging Face and a post by the commentator Kimmonismus summarising the company's results. The Hugging Face page lists it as an improved version of the 3.08-billion-parameter TwIL-LM3. Nobody outside webAI has reported testing it yet.
In webAI's own tests, as relayed by Kimmonismus, Pro is roughly level with Alibaba's Qwen3-8B on the headline formal-logic composite while using less than half the parameters. The same post says it leads the public VibeThinker-3B checkpoint on all six formal tasks webAI reports. These are the company's numbers, run on its own benchmark selection, and have not been independently reproduced.
The post says the team specialised IBM's Granite 4.2 through post-training for three jobs: judging whether conclusions follow, translating statements into formal logic, and critiquing proofs. The Hugging Face page describes the earlier TwIL-LM3 as built on SmolLM3 with LoRA fine-tuning, weight interpolation and reinforcement learning, so the base model for Pro rests on the post alone.
For local use, Kimmonismus says the Q4 GGUF file is 2.09 GiB and runs through llama.cpp. The Hugging Face page lists the earlier model's Q4_K_M file at 1.78 GiB. It carries webAI's non-commercial licence, and webAI warns that it is specialised for formal logic, is not a general assistant and has no safety or preference tuning beyond its base model.