Use open-weight models for narrow jobs like classification and PII detection, and rent them from a provider instead of self hosting

Open-weight models are not a replacement for frontier models. They are a cheaper tool for the boring, well defined steps that make up most of a workflow.

What happened

In a VentureBeat guest post, Sean Falconer, VP of product for AI at Confluent, argues that open versus closed models is not a winner takes all choice. He says the bigger risk is getting so locked into one provider that switching becomes too costly. He points to an Andreessen Horowitz survey where 37% of enterprise respondents were using five or more models.

His view on where open-weight models fit: narrow, well defined workloads like embeddings, classification and detecting personally identifiable information. Frontier models charge for flexibility you may not need there.

On running them, he says most companies should use an inference provider such as Fireworks or Together AI. Self hosting makes sense only when AI is core to the business and you have the scale and expertise. He also warns that open-weight is not the same as open source, and that with open models, the safety work shifts to your own team. Finally, he notes a customer facing app often cannot absorb extra latency, while an internal task like summarizing support tickets usually can.

My take

This matches how I split models in client builds. A typical lead or ticket workflow has five or six AI touches, and maybe one needs a frontier model.

1. Tagging, routing, language detection and PII masking are classification jobs. A small hosted model handles them for a fraction of the cost.
2. PII detection is a good reason to look at open models: with the right provider or your own cloud, you choose where the data is processed, which matters for clients in regulated fields.
3. Keep the frontier model for the reply the customer reads or the decision that needs reasoning.
4. Build the workflow so the model is a setting, not hard coded. When prices drop, you switch in minutes.

Mixing models is not complexity for its own sake. It is how you keep the monthly bill flat as volume grows.