Distillation started as a compression trick, not a philosophy. Researchers used it to squeeze a large, expensive model into a smaller one that ran faster and cheaper, trading some accuracy for deployability. Sergio Paniego, writing on X, traces how that narrow technique became the standard method for building competitive open models.
The shift happened because distillation stopped being only about size. Instead of mimicking outputs purely to compress a model, labs began using a frontier model’s outputs, its reasoning traces, its instruction-following patterns, as training signal for a separate, often much cheaper model. Paniego argues this is the real turning point: distillation moved from compression to capability transfer. The student model does not just get smaller. It gets smarter, faster than it would through its own training run.
That is the mechanism behind the DeepSeek, Qwen, and GLM open-model families. Rather than fund a full pretraining run and then post-train it from nothing, a lab can generate large volumes of high-quality reasoning data by querying a frontier system, then fine-tune a smaller model on that data. The frontier lab did the expensive part: discovering what good reasoning looks like. The distilling lab inherits it at a fraction of the cost.
That inheritance is why distillation became a legal flashpoint. If a model’s outputs are treated as a trained artifact, generating millions of completions from it and training a competitor on those completions looks like copying the artifact without a license. If a model’s outputs are treated as ordinary text, no different from any other publicly generated content, then training on them looks like the same web-scale learning every frontier lab relied on to build its own models. Neither framing has been settled by a court ruling anyone can point to. The dispute persists because both readings are internally consistent and mutually exclusive.
Frontier labs are not waiting for that dispute to resolve. Restricting API access, rate-limiting reasoning traces, and suspending accounts that query at distillation scale are defensive moves aimed at the labs downstream that would otherwise distill a model for a fraction of its training cost. The restriction itself is evidence of how seriously frontier labs treat distillation as a competitive threat, not only a legal one.
For builders, the practical read is this: distillation’s maturation explains why open-weight models closed so much ground on instruction-following and reasoning benchmarks without matching frontier-scale compute budgets. That gap closing is not purely a research story. It is a direct transfer of value from the labs that pay for pretraining to the labs that pay for API calls and post-training compute.
Every new access restriction from a frontier lab should be read first as an anti-distillation measure, and only second as a safety or abuse policy. The economics of who gets to inherit frontier reasoning for cents on the dollar are the real stakes.
Teams evaluating an open-weight model’s reasoning quality should ask which frontier model’s outputs trained it. That lineage now predicts capability better than parameter count does.
Based on commentary from Sergio Paniego, writing on X in July 2026.
