open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → cut development cost, not deployment cost. If you use a model off the shelf with little internal iteration, a closed hosted API is almost always cheaper, because its price comes from economies of scale across everyone's usage.
Open-source software gets cheaper as it spreads because users fix the bugs, so deployment at scale is the cheapest outcome. Open models break that loop: users cannot feed improvements back into the weights, so nearly all the cost stays with the model developer.
The real gap in the open stack is tooling, not weights. There is still no truly open recipe for large-scale reinforcement learning on mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → models, because companies keep forking open tools into internal-only versions sold as enterprise features.
Closed labs read the open technical reports too, but sitting a few months ahead on the development tree means shared findings are worth less to them. The compounding mostly benefits whoever is behind, which pulls open labs onto a common performance curve.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
Why it matters
It reframes the open-vs-closed model debate around R&D economics — arguing open models cost more off-the-shelf but compound advantages for teams doing internal development — which helps engineers reason about whether to build on Chinese open-first labs versus closed frontier APIs.
Key quotes
“Most of the compute to build a leading frontier model comes from R&D costs, rather than the compute to train the final, big model end-to-end.”
“If someone is going to be just using AI off-the-shelf with minimal iteration or internal development, using open models will almost always be more expensive.”
“The more open the stack is, and the more information is shared, the more costs are reduced in future iterations.”
