Small models move the frontier: 70B-class quality on 3 billion parameters

The model power map was redrawn completely in two years. Models of 3–8B parameters like Phi, Gemma and the small Qwens now pass their 50–70B predecessors on text benchmarks. No magic: distillation (training on big-model outputs) plus small, polished training data. The result: frontier-grade quality on a fraction of the memory and energy.
For developers this reopens three decisions. First, routine work — ticket triage, summarization, RAG rerankers, templated answers — never needed a frontier model; a small model on the same GPU answers cheaper and dozens of times faster. Second, on-device runs (laptop, phone, edge box) moved from dream to practice: a quantized 3B fits in 8 GB of RAM. Third, the big models still own their ground: deep reasoning, mathematics and multi-document analysis.
The winning 2026 pattern is routing: small model by default, frontier model only for the hard task. That is a bill that stops growing with volume. The site’s model table exists for exactly this — line up parameters, price and quality, and start your team with "what is the smallest model that works?", not "what is the best model?".