Eighteen months ago, the consensus was that open-weight models would always trail the frontier by a year or more. That gap has been closing fast, and the reasons are more interesting than “more compute.”
What’s actually changed
Three things converged: better synthetic data pipelines, more efficient architectures reaching parity with brute-force scaling, and — critically — frontier labs open-sourcing older-generation weights that are still extremely capable.
- Distillation techniques now recover 90%+ of a teacher model’s benchmark performance at a fraction of the parameter count.
- Post-training (RLHF, DPO variants) has become commoditized enough that small teams can meaningfully close quality gaps without frontier-lab budgets.
- Inference costs for open models dropped an order of magnitude on specialized hardware, making self-hosting viable for real products.
What it means for builders
If you’re choosing a model today, “open vs. closed” is no longer the first question — task fit, latency, and cost usually matter more. That’s a genuine shift from even a year ago, when closed frontier models were the default unless you had a specific reason not to.
The teams winning right now aren’t the ones with the biggest model. They’re the ones who picked the smallest model that reliably does the job.