[Study/Models] Flint: Compressing Reasoning Without Breaking It

TLDR: I trained Qwen3.5-4B and gemma-4-12b on self-distilled, compressed reasoning traces; compression was section-aware (compute and verification spans remain, fillers/narration/transitions get dropped/compressed). The models match or beat the originals, often by a large margin, while using 2

中文全文翻译正在补齐，可先查看下方英文原文。

TLDR: I trained Qwen3.5-4B and gemma-4-12b on self-distilled, compressed reasoning traces; compression was section-aware (compute and verification spans remain, fillers/narration/transitions get dropped/compressed). The models match or beat the originals, often by a large margin, while using 2