3.66B TwIL-LM3-Pro Matches Qwen3-8B on Logic Benchmarks and Runs on CPU

CommonMinimum587 · reddit · 2026-10-13

webAI released TwIL-LM3-Pro, a 3.66B model built on Granite 4.2 3B and post-trained for formal logic (LoRA SFT + checkpoint merging + RL against a programmatic verifier). It scores 55.4 logic composite — beating its 43.1 base and tying Qwen3-8B (53.4, attributed to sampling noise) — with 41% vs 17% on strict multiple-choice logic. Weaknesses: gpt-oss-120b still leads on rule induction and Lean formalization, general average is 79.0, and it thinks 1,900 tokens per answer. Notably it runs 10.5 tok/s CPU-only on a Xeon; Q4KM GGUF is 2.09 GiB with Ollama/llama.cpp support (temp=0, 2048+ token budget to match published scores).

Original post →

More from Infra

• OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
• Jensen Huang lays out NVIDIA's full Vera Rubin agentic system: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
• Jensen Huang details Vera Rubin agentic stack: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
• ONNX to Transformers.js: running models directly in your browser — JosephJacks_ · 2026-10-23
• MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
• "We bet on Cloudflare and it paid off": R2 and D1 at billion scale — dinasaur_404 · 2026-10-23

Infra channel →