Choose your Model for Outcome-Maxxing
Last week, Replit joined NVIDIA, Microsoft, Meta, and dozens of other technology companies in signing a letter supporting open-weights models.
Its central argument was simple: the future of AI will not be determined by one frontier model. Rather, it will be determined by whether we build a broad ecosystem that can match the right model to the right job at the right cost.
Today, we’re putting that principle into our product.
We’re introducing model choice on Replit, ranging from the open-weight Kimi K3 to a careful selection of closed-weight models that we evaluated extensively with ViBench.
For most of this AI era, model choice barely mattered. Capable intelligence was scarce, and the best strategy was usually to leverage the strongest frontier model available. That is why Replit Agent allowed you to select different price points, but didn't expose the model choice in exchange of better performance and reliability.
The status quo is changing quickly, though. Models are proliferating. They vary in capability, speed, cost, openness, and domain expertise. A model that is best at planning may not be best at editing code. A model that excels on a difficult reasoning problem may be wasteful for a routine task. The future of agents therefore cannot be built around a permanent dependence on a few frontier models.
Replit’s job is not simply to pass a prompt to a model and return its tokens. It is to find the best intelligence for the task, give it the right context and tools, evaluate what it produces, refine the result, and turn it into working software.
Open-weight models are important to that future. They expand the supply of intelligence, increase competition, reduce dependence on any single provider, and allow models to become more specialized and adaptable.
They also change the economics of agents. An agent may generate and process enormous numbers of tokens while completing a task. It makes little sense to pay frontier-model prices for every step. An optimal agent will always use frontier intelligence where it is necessary, and efficient intelligence everywhere else.
From today, users can express their model preference directly. Over time, we expect Replit to make all the right decisions dynamically: selecting different models for different parts of a task and searching for the best outcome across them.
Frontier Labs are building increasingly capable sources of intelligence. As an Agent Lab, our work is to take intelligence from across the ecosystem and make it useful. Open-weight models provide more sources of intelligence to choose from, hence ensuring that no single company determines what can be built with it.