Twelve months ago, the conventional wisdom was clear: open source models were interesting for research and hobbyists, but couldn’t touch the frontier. GPT-4 and Claude were simply in another tier.
That framing has aged poorly.
What changed
The training data quality story has been rewritten. Synthetic data generation — using frontier models to produce high-quality training examples at scale — has dramatically raised the ceiling for what’s possible with smaller architectures. You don’t need a hundred billion parameters if your training signal is clean enough.
Quantization has improved to the point where a model that technically requires 80GB of VRAM can run meaningfully on a gaming laptop with 12GB. The degradation is real but acceptable for many use cases.
And the community momentum has been extraordinary. The rate of iteration on open weights models has outpaced the closed labs in deployment speed, if not in raw capability.
What this means practically
For individuals: running a capable model locally is now within reach for anyone with a reasonably modern computer. The privacy implications alone make this worth paying attention to.
For businesses: the build-vs-buy calculus is shifting. Fine-tuning an open model on proprietary data and running it in your own infrastructure is increasingly viable, and the data sovereignty arguments are compelling.
For the labs: the moat has narrowed. It’s not gone — frontier research is genuinely hard and expensive — but the gap between what’s in the cloud and what’s on your desk has shrunk faster than the business models assumed.