Replit starts model choice rollout with Moonshot AI's Kimi K3
Amjad Masad is making model routing a product feature as Replit's usage-based economics put cost beside capability.
By Ryan Merket · Published
Why it matters
Replit is turning model routing into a cost and product-control layer, using open weights to reduce dependence on the closed labs that supply AI coding platforms.
Replit began rolling out model choice on July 28th, starting with Moonshot AI's Kimi K3, founder and CEO Amjad Masad (@amasad) said in a post on X.
The launch puts an open-weight model directly into Replit's product strategy days after Replit joined Microsoft, Nvidia, OpenAI, Google, Meta and other technology companies in backing a policy statement against premature restrictions on downloadable AI models. "The future is the right model for the right job, at the right cost," Masad wrote.
For Masad, the move extends a career built around making developer infrastructure easier to reach. Before starting Replit in 2016, he was Codecademy's first employee and helped build its browser-based programming environment. He later started Facebook's JavaScript infrastructure team, working on developer tools including Babel, Jest and React Native infrastructure, according to his personal biography.
Replit puts model economics inside the product
Masad described Kimi K3 as the starting point for model choice across Replit. The announcement follows Replit's November 2025 release of AI Integrations, which lets customers connect applications built on Replit to more than 300 third-party models without separately managing API keys or provider accounts.
That earlier system focused on the models inside a customer's application. The Kimi K3 rollout brings the model-selection question closer to Replit's own app-building workflow. Fireworks AI, an inference provider serving Kimi K3, published a comment from Replit President Michele Catasta saying the model had been tested on Replit Design and delivered strong interface-generation results at a lower cost than the frontier models Replit typically uses.
The distinction matters because Replit charges for AI features according to usage and effort. Its billing documentation says third-party API calls made by Agent are deducted from customer credits at the provider's public API rate. Replit's current pricing page includes $25 in monthly credits with its Core plan and $100 with Pro.
Giving customers a model menu can make those costs easier to control. It also gives Replit room to route work toward cheaper models when a closed model's additional capability does not justify its price. Masad's wording makes cost an explicit part of the selection logic, alongside task fit.
Replit is making that shift after announcing a $400 million financing at a $9 billion valuation in March. Georgian led the Series D, with participation from G Squared, Prysm Capital, Coatue, Andreessen Horowitz, Craft Ventures, Y Combinator and other investors. The financing gives Replit capital to expand Agent while competition intensifies among browser-based builders, coding agents and AI-native development environments.
Kimi K3 gives Replit an open-weight alternative
Moonshot describes Kimi K3 as a 2.8 trillion-parameter mixture-of-experts model with 104 billion parameters activated for each token. It supports images, a context window of roughly 1 million tokens and long-running coding and knowledge-work tasks.
Moonshot released the model's weights under its own Kimi K3 license, allowing developers to download, deploy, modify and fine-tune the system subject to the license's conditions. That gives infrastructure providers and larger customers deployment options unavailable with models offered solely through closed APIs.
The July 24th open-weight statement signed by Replit argued that downloadable models can lower costs, reduce dependence on individual vendors and let customers run AI on infrastructure they control. It also acknowledged that released weights cannot easily be recalled and that modified versions can be difficult to trace.
Replit now has a commercial reason to test that policy position in production. Model choice can reduce reliance on any single lab, create pricing pressure among providers and give Replit another control over the cost of running Agent. Kimi K3 is the first test of whether open-weight models can compete for work inside a heavily funded consumer and enterprise software builder, rather than remaining an option developers must host and integrate themselves.