Your model is not the part of your agent setup you actually own. Your playbook is.

Mistral Large 4 has been live as an API preview since 6 October. About a trillion parameters, roughly 50 billion active, and Mistral says the open weights land at the end of this month.

This is where it gets interesting. If your agents run on written skills, Standard Operating Procedures (SOPs) and living docs, a big open-weight model is a real option, not a rebuild.

Swap the engine, keep the playbook.

I run a small fleet of agents this way. The skills, the handoffs and the review gates are all written down and backed up in GitHub. None of that is tied to one model. When something new shows up, I point the same procedure at it and see what breaks.

Teams that kept everything in long prompts and tribal knowledge will have to rebuild before they can even test it. Teams that wrote things down get to try it the week it lands.

One honest caveat. Open weights are still a promise. No open-weight licence is published and there are no weights yet, so self-hosting is not on the table today. Before you route real work to it, put the licence name, the weights date and the hardware footprint into your procedure, the same as any other dependency.

End of October is when the real test begins. Get your playbooks ready.