Learn
Practical guides to fine-tuning, distillation, and deploying small language models.
Can a Small Model Beat Its Teacher?
Yes, on narrow tasks — a 0.6B student beat a 120B teacher by 29 points and a 350M student beat the same teacher on two of three tool-calling benchmarks. The measured figures and why it happens.
Cost Per Million Tokens: Local vs API
A side-by-side of self-hosted small-model rates against per-token API pricing, and why the two figures are not directly comparable even when both are quoted per million tokens.
Does Base Model Accuracy Predict Fine-Tuned Performance?
Across four mid-size models on seven tasks the base ranking inverted after fine-tuning — the worst zero-shot model tied for best once trained. Here are the ranks, the method, and what still transfers.
GPU Requirements and Cost for SLM Inference
Weight-size arithmetic for the distil labs student catalog alongside one measured serving benchmark — 222 RPS and 7.6 GiB for a 4B model — and how a GPU-hour bill becomes a per-request price.
How Long Does Fine-Tuning Take?
Published distil labs runs land between roughly six and twelve hours end to end. Here is what those figures cover, which stage eats the clock, and what moves it.
How Many Examples Do You Actually Need to Fine-Tune a Model?
Twenty labelled examples is the documented floor and 20 to 100 the usual range — here are the published results behind those figures and what moves the number in either direction.
How Much Can You Save Replacing an LLM API?
Knowunity cut inference cost by roughly 68% on selected tasks while accuracy rose from 81% to 93%. Here is that measurement, the wider range published alongside it, and what puts you at either end.
How Much Traffic Do You Need Before Traces Are Useful?
There is no published traffic threshold for trace-based training. This works out what the documented processing pipeline — dedup, seed splitting, relevance filtering, relabelling — implies about volume instead.
How Much VRAM Does a 1B, 3B or 8B Model Need?
Working forwards from a chosen model size to the card that can serve it, including the KV cache term computed from published attention configs — the part that decides whether a model that loads will still run at your context length.