When the frontier is downloadable, what exactly are you paying for?

Open weights used to trail the frontier by a year and apologise for it. This week they arrived at the top of the range, with a self-hostable sibling behind them, and the closed tier's pricing power became a live question.

Alibaba is following its 2.4-trillion-parameter flagship with open weights and a 27-billion-parameter variant — one model to rent, one to run on hardware you already own, from the same family, weeks apart. That pairing is the whole story.

The 27B is the one that moves markets

Almost nobody self-hosts a 2.4-trillion-parameter model; the flagship is a rental. But a 27B variant carrying most of the family's behaviour is deployable inside a bank, a hospital, a defence contractor — anywhere the binding constraint is governance rather than capability. The frontier number wins headlines. The distilled number wins procurement.

The stack already assumes you exist

You can see the assumption hardening elsewhere: the 2026 coding-agent rankings now list an open-weight model as the self-hosting default and a provider-agnostic harness alongside it. When tooling is designed model-neutral, the model becomes a component — and components compete on price and licence, not mystique.

What the premium has to become

None of this means closed models lose. It means the argument for them has to change. Reliability, deprecation guarantees, indemnity, support, integration depth, uptime — the boring commitments enterprise software has always charged for. What no longer works is charging a premium for capability that has an open equivalent shipping next week.

The open tier stopped asking to be taken seriously somewhere around the point it started setting the price.

TechBriefly — Alibaba unveils open-source AI model Qwen3.8-Max → · Levelop — Best AI coding agents in 2026 →