There’s a quiet assumption baked into most ecommerce AI conversations right now — that you need a frontier model to do anything useful. That if you’re not paying for enterprise-tier API access, you’re not really in the game.
The hackathon results coming out of the open-weights community are starting to poke holes in that. Multi-agent shopping workflows — cart logic, product search, intent parsing, recommendation chains — running on 3B parameter models. Not degraded. Not “good enough for a demo.” Actually handling the token load that real transactions generate.
What that means operationally is that the infrastructure cost conversation changes shape entirely. We’re talking commodity hardware, open weights, no per-token billing on a subscription that scales against you every time a campaign lands. The gap between what a small operator can ship and what a funded team can ship just got narrower.
I’ve watched merchants get quoted serious monthly figures for AI integrations that, under the hood, were doing surprisingly simple retrieval and ranking tasks. The model was almost incidental. The cost was in the API wrapper and the margin the vendor was taking on top of it. That math looks different when the model runs locally.
The part I’m still turning over is the maintenance side. Open weights means you own the update cycle. That’s not free — it’s just a different kind of expensive.