You are probably skimming for a yes or no on "should we leave the API." I will not give you one in the first line. Hugging Face published State of Open Models: Summer 2026 Observations on 14 Aug. Their claim is that specialization, inference cost, and tooling now decide adoption more than a raw parameter count. Mistral, three days earlier, talked about in-region inference and European capacity for what they call sovereign AI. Same month, two vendors, one budget meeting.
What the meeting is actually about
Data residency. Exit cost. Whether legal will accept a US endpoint for tickets that contain customer PII. Whether your eval set still passes on a 27B open model or you are lying to yourself with a chat demo.
IT consulting services here means we sit in that meeting with a matrix, not a slide that says "AI strategy." Closed API vs open weights vs local runtime. Latency. License. Fine-tune ownership. Who gets paged when the model starts refusing the workflow you sold.
If you already decided the model lives on your metal, read thirty billion on the desk, three billion awake. That post is the Ollama box. This one is the purchase order.
What we write down
A one-page decision: task, data class, region, fallback, kill switch. If the task is "summarize a public handbook," open weights are often enough. If the task is "decide credit," I want evals you can rerun in six months, not a leaderboard screenshot from August.
Mistral's regional pitch is a product. Hugging Face's roundup is a weather report. Neither is your architecture. We will argue about where the weights sit and who can export the fine-tune. Request a quote if that argument is currently a Slack thread with three conflicting screenshots.
I skip the section where every article says the landscape is evolving. You already know it moved. The unpaid work is picking a default and dating it.
Sources
- State of Open Models: Summer 2026 Observations — Hugging Face, 14 Aug 2026
- In-region inference, open models, and new European infrastructure for sovereign AI — Mistral AI, 11 Aug 2026