Open weights with an invoice

A 2.4-trillion-parameter model you can download, under a licence that wants a share of what you earn with it. That is a third category, and it needs its own name.

Alibaba put a 2.4T sparse MoE on Hugging Face, the first Max-tier Qwen ever downloadable — under a revenue-share licence.

Three things are now called open weights

Apache 2.0 and MIT: take it, ship it, no conversation. Community licences with acceptable-use appendices and user-count tripwires: take it, and have a conversation with your lawyer. Revenue share: take it, and have a conversation with their finance team.

Those are not variations on a theme. They are different products, and the industry has spent a year using one phrase for all three.

Why it might still be the right trade

Self-hosting real weights buys things no API offers: inspection, modification, data residency, and immunity from a vendor deprecating your model. For an organisation that needs those, paying a percentage may beat paying per token — and unlike an API, the terms are visible up front.

The questions that decide it are the boring ones. What counts as revenue, what triggers the obligation, is internal use captured, and what happens when the model is three layers inside a product.

The asterisks

The checkpoint is text-only — the vision and 1M-context capabilities of the hosted model are absent. And the 27B promised for the same week has no repository, licence or date, which is the model most people could actually have run.

A frontier-scale model you cannot serve and a small model that has not shipped is a generous release and a thin one at the same time. Both readings are fair.

explainX.ai — Qwen3.8-Max Open Weights Are Live (August 2026) → · OrcaRouter — Qwen3.8-27B Open Weights: What We Know Before the Drop → · CNBC — Meta to open source its most powerful AI model →