webAI 发布了 TwIL LM3 Pro，一个基于 IBM Granite 4.2 3B、仅针对形式逻辑（如前提蕴含判断、规则归纳、Lean 证明批判）微调的 3.66B 模型，在逻辑复合基准上以一半以内的参数量追平 Qwen3-8B，并提供可在 CPU/4GB VRAM 上运行的 GGUF 量化版本。
webAI released TwIL LM3 Pro on September 30. I haven't seen it posted here yet, so I went through the model card, and their comparison chart is attached. It's a 3.66B model built on IBM's Granite 4.2 3B and tuned only for formal logic. That means things like checking whether a conclusion follows from its premises, rule induction, entailment and critiquing Lean proofs. They post trained it with LoRA SFT, checkpoint merging and RL against a programmatic verifier. The same recipe lifted VibeThinker-3B from 37.4 to 54.1 . On their logic composite it scores 55.4, against 43.1 for the Granite base, 42.2 for the original TwIL-LM3, 41.2 for VibeThinker-3B and 53.4 for Qwen3-8B. The card itself calls the Qwen3-8B gap sampling noise, so that's a tie at less than half the size. Where it clearly leads is strict multiple choice logic, at 41% against 17% for the next model, and BBH logic at 95.4%. They also publish where it loses. gpt oss 120b is still ahead on rule induction, entailment and Lean formalization. On general benchmarks it averages 79.0, against 81.0 for VibeThinker-3B and 84.9 for Qwen3-8B. It also thinks long on logic tasks, around 1,900 tokens per answer, and a quarter of answers hit the length cap llama.cpp If you want to try it, the Q4_K_M GGUF is 2.09 GiB and runs on CPU or 4 GB of VRAM, with Q5, Q6 and Q8 builds up to 3.63 GiB. It runs in Ollama, LM Studio and llama.cpp straight from the Hugging Face page, and the model card lists the exact commands. Keep the temperature at 0 to match their numbers, and give it at least 2048 tokens so the thinking doesn't get cut off. Their scores are on BF16 weights and webAI hasn't benchmarked Q4 yet. The license is non commercial want to test it on policy rules with exceptions, contract conditions, and as a checker step in an agent pipeline before anything acts. If you've run it, how did Q4 hold up against their numbers, and did the long thinking get in the way? Model card and full eval tables: https://huggingface.co/webAI-Official/TwIL-LM3-Pro
IBM has launched Granite 4.2, a family of open-weight large language models with variants up to 30B parameters, featuring a 128,000-token context window and agentic reinforcement learning for enhanced capabilities.
Granite 4.2 is IBM's new family of reasoning LLMs available in 3B, 8B, and 30B sizes, featuring thinking modes, tool calling, and trained with a multi-stage reinforcement learning pipeline under the Apache 2.0 license.
This article details the technical architecture and training pipeline of IBM's Granite 4.1 LLMs, covering pre-training, SFT, and RL stages. It highlights that the 8B dense model outperforms larger MoE counterparts and notes the release under Apache 2.0 license.
webAI released TwIL-LM3-Pro, a 3.66B open-source reasoning model that surpasses VibeThinker-3B, Qwen3.5-4B, and LFM2.5-8B-A1B on formal logic benchmarks and runs locally via llama.cpp in a ~2 GiB Q4 build.
TwIL-LM2, a specialized 1.7B model fine-tuned for formal logic translation, outperforms larger generalist models like Qwen3-8B and Gemma-4-26B on strict scoring benchmarks, highlighting the potential of narrow AI specialists for efficient reasoning.