Worker

Model distillation

A distilled model is a small model trained to imitate a much larger one. The large model is the teacher; the small model is the student, taught to reproduce the teacher's answers. The result is a model with far fewer parameters that keeps a surprising amount of the bigger model's quality — especially its reasoning. In Pendra's catalogue, the DeepSeek-R1-Distill family is the clearest example: 1.5B–70B students distilled from the full 671B DeepSeek-R1.

Pendra flags these models with a Distilled badge in the console — on the Models grid and on each model's page — so you can spot them at a glance. It's a fact about where the model came from, not a capability like chat or vision.

Why it matters when you pick a model

A distilled 8B does not behave like an 8B trained from scratch. Because it learned from a much stronger teacher, it tends to punch above its size on the kinds of tasks the teacher was good at — for reasoning models like DeepSeek-R1-Distill, that means step-by-step problem solving. You get a lot of that reasoning quality at a fraction of the memory, cost, and latency of running the full model.

The trade-off

Distilled modelFull-size model
Much smaller — runs on modest GPUs and CPUsNeeds far more memory and disk
Faster, cheaper, more concurrent requestsSlower, more expensive to serve
Keeps much of the teacher's strengthsThe full range of the original model
Narrower — strongest on what it was distilled forMore general across tasks

When to pick a distilled model

  • You want strong reasoning on modest hardware. A distilled reasoning model can give you chain-of-thought quality that would otherwise need a much larger model and a much bigger GPU.
  • Cost and speed matter. Fewer parameters means faster responses and more requests in parallel for the same hardware.
  • Your workload matches what it was distilled for. Distilled models are at their best on the teacher's strengths; for very different tasks, a general model of similar size may be the safer pick.

Distillation is one of several levers on model choice — see also choosing a model size, model architecture, and choosing a quantization. The catalogue lists distilled and full-size models side by side; pick the one that fits your task and your worker's memory.