Best open-weight AI models to self-host

What to weigh for self-hosting

Open weights change the question from 'what does it cost' to 'what does it cost to run': you pay for hardware and operations, and the data never leaves your network. Only models whose weights are published qualify here; the list price shown is what hosted providers charge for the same weights, a useful floor for what self-hosting has to beat.

The strongest open-weight models now sit close to the closed frontier on the intelligence index, and they are the only answer for a data boundary a provider's terms cannot cross. The gap that remains is in the hardest reasoning and agentic requests - which is exactly what a router is for.

The usual deployment keeps a self-hosted open-weight model as the default and lets the router escalate to a hosted model only for requests that do not carry restricted data and do need the extra quality. A data-boundary rule pins the sensitive requests on-premises; the rest float to whichever target wins on price and quality. Because the self-hosted target has a real cost too - a GPU-hour divided by the tokens it served - declare it in the catalogue, and the router's savings figures compare like with like instead of treating your own hardware as free.