Hugging Face engineer Sergio Paniego sorts the past year's distillation usage at the frontier labs into three main lines. The first: the classic "big teacher with small student". Gemma 3/4 and DeepSeek-R1-Distill take this path, compressing the large teacher's capability on the next-token distribution or generated text down to a small-sized student. The second: using distillation to merge multiple RL experts into one student — this is the direction that everyone has truly converged on this year. DeepSeek-V4 trains domain experts for math, code, and Agent separately, then merges them back into a single model via on-policy distillation; MiMo-V2-Flash names it MOPD; NVIDIA Nemotron 3 Ultra scales up to a dozen teachers; GLM-5 uses it to recover capabilities forgotten late in RL. Qwen3 gives the key number: this path costs only 1/10 of pure RL GPU-hours, and the effect is actually better. The third: self-distillation. Cursor Composer 2.5 uses "itself with hints" to align with "itself without hints"; Thinking Machines uses "previous-version self" to distill back to "current-version self", directly solving the continual-learning problem. The three lines are essentially the same teacher-student regression at different scales — distillation is evolving from a "compression tool" to a "training paradigm".