Frontier models get the headlines. Seven-billion-parameter fine-tunes do most of the actual production work in the systems we deploy. Where small models are quietly winning in 2026 — and where they still lose.
Where small wins
- Classification. Intent, sentiment, topic, routing. A 7B fine-tune matches a frontier model at 1/30th the cost.
- Extraction. Entities, dates, amounts. Same story.
- Reranking. A small cross-encoder is the right tool. Don't ask a frontier model to do this.
- VAD and intent classifiers in voice. Latency-critical, run on-device or near-edge.
- Guardrails. Prompt-injection detection, PII redaction, content filtering. Small models, fast, cheap, easy to keep current.
- Summarization at scale. Where you have millions of items, mid-tier or small fine-tunes amortize.
Where small loses
- Open-ended reasoning. Frontier models pull ahead.
- Long-context understanding. Small models have shorter context and worse coherence over it.
- Tool-calling under uncertainty. Frontier models are noticeably better at deciding when to call which tool.
- Highly-specialized domains where the frontier has been trained on more data than you can fine-tune on.
The two-tier pattern
Almost every production system we deploy now uses two tiers. A small model handles 80%+ of requests cheaply. A larger model handles the cases the small model isn't sure about. The router is itself a small classifier. The cost curve is dramatic — often 5–10× cheaper than running everything through the larger model — with no measurable quality drop.
What changed in 2026
Open-weight 7B–13B models in 2026 are roughly where 70B models were in 2024. The gap to the frontier is real but narrower than the gap was. The economics of small fine-tunes have improved dramatically: a useful fine-tune now costs hundreds of dollars, not tens of thousands.
Don't reach for the frontier model first. Reach for the smallest model that hits your quality bar. You will be surprised how often that's enough.