webAI's TwIL-LM3-Pro Brings Formal Logic Reasoning to 4 GB GPUs
webAI's 3.66B Granite-based model specializes in formal logic, running locally in 2 GB while beating much larger reasoning models on logic benchmarks.
Takeaways
- webAI released TwIL-LM3-Pro, a 3.66B formal-logic model fine-tuned from IBM Granite 4.2-3B.
- Macro gate on in-domain logic rose from 0.4313 to 0.5539 (+28% relative) over the base model.
- Beats VibeThinker-3B, Qwen3.5-4B, and LFM2.5-8B-A1B on formal logic by double-digit margins per webAI's harness.
- Five-stage pipeline: LoRA SFT, multipath distillation, checkpoint fusion, WiSE-FT merge, and entropy-weighted GRPO (MGPO).
- Runs locally via Q4_K_M GGUF at 2.09 GiB on CPU or 4 GB VRAM; verbose at 1,902 tokens per answer.
- Shipped under webAI Non-Commercial License 1.0 — research only, no commercial deployment without separate agreement.
webAI has released TwIL-LM3-Pro, a 3.66-billion-parameter model tuned for symbolic reasoning. Derived from Granite 4.2-3B, it fits in a 2.09 GiB Q4_K_M file and can run through llama.cpp on a CPU or a GPU with 4 GB of VRAM. That footprint gives researchers a local option for logic experiments, though internal benchmarks and a noncommercial license limit its immediate use in products.
A compact logic specialist
The release targets first-order-logic translation, entailment labeling, semantic parsing and Lean proof critique. These tasks cover converting prose into symbolic statements, deciding whether premises support a conclusion, mapping language into formal structures and identifying problems in theorem-prover output.
General conversation and broad instruction following fall outside the model’s stated scope. webAI also added no safety or preference tuning beyond the training already present in Granite 4.2, so deployments would need their own evaluation and control layers.
Logic gains, held-out score steady
webAI’s evaluation reports a substantial gain on its in-domain formal-logic track, called Track A, while performance across a held-out 10-dataset suite changes little:
| Metric | Granite 4.2 base | TwIL-LM3-Pro | Change |
|---|---|---|---|
| Track A macro gate | 0.4313 | 0.5539 | +0.1226, or +28.4% relative |
| Held-out macro | 0.7942 | 0.7901 | -0.0041, or -0.5% relative |
The small held-out decline matches the goal of the model-merging stage: preserve most of the base model’s broader capability while strengthening formal-logic performance. All reported results come from webAI’s evaluation harness and have not been independently replicated.