The five drivers

Why did the frontier come back for the small?

From 2020–2023 the rule was obvious: bigger is better. Every step up the scaling ladder — GPT-2 → GPT-3 → GPT-4 — revealed capabilities the previous step didn't have. So why, in 2024–2026, did the frontier labs all release sub-5B-parameter variants (Phi-4-mini, Llama 3.2, Qwen3, Gemma 3) as first-class citizens of their model families?

Five reasons, none of them “small is cute.” Flip through them on the right — each is a distinct force, and understanding them individually is how you pick the right model for a given job in the rest of the curriculum.

None of these are small-is-better in general

Pay attention to what I'm notclaiming. I'm not saying SLMs outperform LLMs at open-domain reasoning. They don't. I'm not saying you should replace your production GPT-4 with a 3B model and hope for the best. That will disappoint you.

What I'm saying is that each of these five drivers is a real, domain-specific force that can make a smaller model the right answer. In practice most production systems end up with an SLM somewhere in the pipeline (intent classification, tool-call execution, summarization of retrieved passages) even if a larger model handles the free-form generation hop. The next lesson makes that concrete.

A worked example of the five-driver stack in one system. Imagine a voice agent at a mid-market SaaS company in 2026: Deepgram (ASR) → Qwen3-1.7B intent classifier → Phi-4-mini tool-caller → GPT-4o-mini fallback for open-ended answers → ElevenLabs (TTS). Four models in the hot path, and only one of them is over 4B parameters. By the lesson's 2N FLOPs approximation, a 1.7B classifier on co-located commodity GPU silicon is well under the 300 ms first-token budget; if its confidence threshold is tuned so that most user turns clear it, the fallback only fires on the long tail of open-ended questions it can't classify. Every driver you just read is in there: inference economics (the classifier is two orders of magnitude cheaper per token than the fallback), edge-ish deployment (the SLMs colocate with the audio pipeline, eliminating a network hop), latency (first-token under 300 ms because prefill is tiny), specialization (the tool-caller was fine-tuned on the company's own function registry), and velocity (they re-ran the classifier fine-tune in three hours last Tuesday). The frontier model isn't banished — it's rationed.