Deepseek

An experiment in fine-tuning small, local language models (Qwen 1.5B) to reason in a telegraphic, compressed style, saving up to 74% of reasoning tokens and speed up inference without relying on verbose prompts.