Enterprises Shift to Open-Weight AI Models, Putting 85% of Compute Into Custom Training
Updated
Updated · InfoWorld · Aug 11
Enterprises Shift to Open-Weight AI Models, Putting 85% of Compute Into Custom Training
3 articles · Updated · InfoWorld · Aug 11
Summary
Enterprises are moving from one-time model selection toward continual learning loops that use outcome data to decide when cheaper AI models are sufficient and when custom tuning is needed.
That shift reframes cost: a low-token-price model can be more expensive in practice if staff must review, retry or repair weak outputs, making business outcomes—not benchmark scores—the key metric.
Open-weight models are gaining traction because they let companies carry learning forward as models change, turning base models into raw material they can specialize with their own workflow and customer signals.
Cursor’s Composer 2.5 illustrates the approach: it started from the open-weight Kimi K2.5 checkpoint, and Cursor says 85% of compute went into additional training and reinforcement learning.
Fireworks, Together AI and Baseten are all pushing this category, betting enterprises want closed-API convenience while keeping control of the intelligence their proprietary outcome data creates.