News ›
SouverainCode sets its own level
Fast to start, then Strong and Deep when the task gets harder: the agent steps up by itself, tells you, and you keep control with a selector.
Not every task needs the same effort. SouverainCode sets its own level according to the difficulty. Today, Fast, Strong and Deep use the same Mistral model, Ministral 3 8B, hosted in Québec: stepping up a level does not change the model, it uses up your plan’s capacity faster.
What you see. Two settings sit to the left of the writing box, and you can change them with every message. The first is the power: Fast answers in a few seconds, perfect for short questions and small tasks; Strong and Deep are for long tasks or ones that resist, and use up your plan’s capacity faster. The second is the length of the answer: short or long.
What happens by itself. With the DEV plan, you do not have to choose. The agent starts on Fast. When the mission gets longer or the difficulty keeps coming back, it steps up to Strong, then to Deep, and a message tells you (for example “Switching to the Strong model”). It keeps that level until the end of the mission. If a model does not answer, the agent moves on to the next level rather than letting the mission fail.
Depending on your plan. The Discovery and Solo plans work on Fast. When a plan caps the level, the mission carries on at the allowed level and a message tells you so.
Images. Attach an image to a message: the vision model switches on automatically.
It is all explained in the “Models and routing” section of your Settings, and in the tutorial “03 Fast, Strong, Deep” of the SouverainCode training. In the SouverainBox, it is the same routing as online, at your place.
On video: Eight AI models to write, code, reason and search and How long before the answer? What we know (videos in French).