The 2024-2025 explosion of small models was the sleeper story of the AI era. Models under 4B parameters now handle tasks that required 70B in 2023.
Phi-4 (14B) — the reasoning specialist. Math, logic, code. Punches way above its weight. Microsoft shows benchmark scores rivaling GPT-4 on many reasoning tasks.
Gemma 3n (2B/4B) — Google's on-device flagship. Runs on Pixel phones at usable speeds, multimodal (text+vision+audio in one model).
SmolLM3 (3B) — HuggingFace's answer to 'good enough for everyone'. 128K context, instruction-following trained, permissive license.
Qwen3-1.7B/4B — Alibaba's compact family. Multilingual, long-context, cheap.
Llama 3.2 1B/3B — Meta's on-device play. Best tooling ecosystem, permissive license.
What they're actually good at: classification, extraction, summarization, rewriting, chat templates, structured output.
What they're not: open-ended reasoning, deep analysis, creative writing, anything requiring real-world nuance.
The deployment pattern winning: small model for the 80% of traffic, escalate to big model only when the small one is unsure. Confidence-gated routing — the FinOps answer to LLM costs.