The 975-Billion-Parameter Bait

When a US lab ships a 975-billion-parameter model under an open license and tells the public not to run it as-is, the model is not the product. It is the bait. The weights exist to feed a fine-tuning platform, where recurring revenue lives.

This is the same shape as "free" games and "free" cloud tiers: a zero-priced first step whose purpose is to convert curious developers into paying customers of something adjacent. Simon Willison's post on Inkling, the release from Mira Murati's Thinking Machines Lab, hands the reader the load-bearing line. The lab calls Inkling "not the strongest overall model available today, open or closed," then points at its own Tinker platform as the actual destination. The thin model card and Training Data Documentation stop looking like an oversight once that posture is named: a base for fine-tuning, not a flagship to benchmark head to head.

The repeatable mechanism is straightforward. A credible open-weights release earns developer attention and trust. That attention lands on the lab's own fine-tuning service, where the model can be customized for paying customers. The weights earn the funnel; the platform captures the margin. Rivals without an adjacent service still have to ship strong general-purpose weights to win mindshare, since the most capable open releases of the past year came from labs whose primary product is the model itself.

The strategic question is no longer "is this the best open model," but "which lab uses open weights as a customer-acquisition funnel, and which uses them as the product." The first group will keep its best weights closer. The second will keep competing on capability.