webAI Releases TwIL LM3 Pro: 3.66B Formal Logic Model Built on Granite 4.2 3B

webAI released TwIL LM3 Pro on September 30, a 3.66B parameter model built on IBM's Granite 4.2 3B and fine-tuned exclusively for formal logic tasks. The model targets capabilities such as checking logical conclusions, rule induction, entailment, and critiquing Lean proofs.

Performance and Benchmarks

The model was post-trained using LoRA SFT, checkpoint merging, and reinforcement learning against a programmatic verifier. On a logic composite benchmark, TwIL LM3 Pro scores 55.4, significantly outperforming the Granite base (43.1) and the original TwIL-LM3 (42.2). It also surpasses VibeThinker-3B (41.2) and matches Qwen3-8B (53.4), with the model card attributing the slight gap to sampling noise.

Key strengths include:

• Strict multiple choice logic: 41% (compared to 17% for the next best model).
• BBH logic: 95.4%.

However, the model trails gpt oss 120b in rule induction, entailment, and Lean formalization. On general benchmarks, it averages 79.0, lower than VibeThinker-3B (81.0) and Qwen3-8B (84.9).

Deployment and Limitations

TwIL LM3 Pro generates long reasoning chains, averaging 1,900 tokens per answer, with 25% of responses hitting the length cap in llama.cpp. Users are advised to set the temperature to 0 and allocate at least 2048 tokens to prevent truncation.

The model is available in various quantizations:

• Q4_K_M GGUF: 2.09 GiB, runs on CPU or 4 GB VRAM.
• Q5, Q6, Q8 builds: Up to 3.63 GiB.

It supports Ollama, LM Studio, and llama.cpp. Note that the published scores are based on BF16 weights, and Q4 performance has not yet been benchmarked by webAI. The license is non-commercial.