Nate@nate_512

Follow

TwIL-LM3-Pro Hugging Face benchmark results

THIS MODEL IS ONLY 2.09 GIB. LET THAT SINK IN What thewebai managed to fit into just 3.6B parameters with TwIL-LM3-Pro is kinda ridiculous: → beats VibeThinker-3B across all 6 formal-logic tasks tested → roughly matches Qwen3-8B on the formal-logic composite → 95.4% on BBH logic → runs locally in ~2GB quantized Tiny model. Serious reasoning. 🤗 Open-weight and fully local (weights in 🧵 ↓)

1h

Nate@nate_512

Follow

TwIL-LM3-Pro Hugging Face benchmark results

THIS MODEL IS ONLY 2.09 GIB. LET THAT SINK IN What thewebai managed to fit into just 3.6B parameters with TwIL-LM3-Pro is kinda ridiculous: → beats VibeThinker-3B across all 6 formal-logic tasks tested → roughly matches Qwen3-8B on the formal-logic composite → 95.4% on BBH logic → runs locally in ~2GB quantized Tiny model. Serious reasoning. 🤗 Open-weight and fully local (weights in 🧵 ↓)

1h