Replit

Replit is a software creation platform whose AI products span code completion, application development, and parallel agent workflows. Its Agent configures development environments, installs dependencies, executes code, and deploys applications. With Agent 4, announced on March 11, 2026, Replit added a design canvas, coordinated parallel tasks, and shared projects containing apps, slide decks, data apps, and animations.[5][8]

These agent products followed Replit’s earlier work as a collaborative browser IDE and its development of Ghostwriter and a code language model. The company was founded in 2016 and joined Y Combinator’s Winter 2018 batch.[3][1][4]

In his 2018 retrospective, founder Amjad Masad recalled leaving his job alongside cofounder Haya, a designer, allocating $20,000 to the startup, and working from their home in Foster City. They initially sold Replit’s code-evaluation infrastructure as an API, but the free service grew faster than revenue. Masad credited Roy Bahat, head of Bloomberg Beta, with backing the company and helping it attract seed financing.[10]

Replit next built a classroom-management product for computer-science teachers. Masad recalled beginning sales in fall 2017 and being on the way to $150,000 in annualized revenue during the first few months. He considered that an encouraging start but judged the school-sales model too slow to support the growth he sought, while the core REPL continued growing.[10]

After three rejected YC applications, Masad recalled receiving a meeting request from Sam Altman, who told him that Paul Graham had discovered Replit on Hacker News. Graham subsequently advised focusing on the core product and user base. Altman invited a late application the night before the Winter 2018 program began; Masad and Haya interviewed the next morning, and partner Adora Cheung called to tell them they had been accepted.[10][3]

Replit’s AI work grew out of its data infrastructure. Reza Shabani, its head of AI in May 2023, explained that when he joined, even a query asking which project had been forked most could time out. His early work modernized pipelines and large-scale processing so data could be queried and prepared in minutes rather than days or longer. That infrastructure supplied groundwork for training and fine-tuning models, alongside search and abuse detection.[1]

Shabani dated the original Ghostwriter code-completion shipment to August 2022. It adapted a smaller Salesforce CodeGen model, using distillation and other optimizations to reduce its footprint and make serving practical. Replit then experimented with fine-tuning for patterns common among its users, including JSX and inline CSS. According to Shabani, the results fell short of the team’s expectations, contributing to the decision to build infrastructure for training models from scratch. By May 2023, Ghostwriter also offered code explanation, generation, transformation, and chat inside the IDE.[1]

replit-code-v1-3b is a 2.7-billion-parameter causal language model focused on code completion. Its training mixture used a subset of Stack Dedup v1.2 covering 20 languages and formats. The dataset contained 175 billion tokens, repeated over three epochs for 525 billion training-token exposures. Training ran on MosaicML with 256 A100-40GB GPUs.[4]

The model used FlashAttention for training and inference, AliBi positional embeddings to support variable context length at inference, the LionW optimizer, and a custom SentencePiece Unigram tokenizer with a 32,768-token vocabulary optimized for code. Replit released the checkpoint and vocabulary under CC BY-SA 4.0 and the Python source files under Apache 2.0, with different licensing obligations for those components.[4]

The larger training run followed months of infrastructure preparation. In his May 2023 interview, Shabani recalled Amjad urging the team to move beyond smaller, single-language experiments and train on all available data before Developer Day. He credited Michele Catasta with proposing multiple epochs, despite the risk of overfitting. Shabani estimated that the push compressed work that might otherwise have taken two or three months into a week. He credited MosaicML with providing GPUs across cloud providers, fault tolerance for hardware failures, and preconfigured training setups that reduced troubleshooting.[1][4]

Replit evaluated models with both HumanEval, which tests generated Python functions against unit tests, and AmjadEval, Amjad’s hands-on assessment. Shabani described models that scored well on HumanEval yet felt unhelpful in use. AmjadEval considered contextual completions, responsiveness, useful code style, and errors that looked plausible. In one interview demonstration, Replit’s model supplied readable constants where CodeGen initialized a value to zero; another example showed Ghostwriter incorrectly assuming that an argument had to be a list. These were qualitative examples rather than a standardized public benchmark. The base model’s card reports a self-reported HumanEval pass@1 result of 0.219.[1][4]

Shabani distinguished the released base model from a version additionally fine-tuned on public, permissively licensed Replit data. He reported improved reasoning despite the presence of beginner code, suggesting that instructional assignments and answers might help explain the result. He also described a future Ghostwriter that could drive the IDE, potentially learning from operational-transform sequences: creating files, editing related files, running code, encountering errors, and responding to them. This was a May 2023 plan; it does not establish the training method used by later Replit Agents.[1]

Replit launched its first-generation Agent in September 2024. Its September 16 announcement said the launch had occurred the previous week and described an application-building workflow that included environment setup, dependency installation, code execution, and deployment. At that stage, Agent was available to Core subscribers in early access and was explicitly described as “alpha” software. Replit reported examples including patient dashboards, a campus parking map, and replacements for workflow-automation tools; Andrew Davison curated a collection of early applications.[5]

On February 25, 2025, Replit announced Agent v2 in partnership with Anthropic’s Claude 3.7 Sonnet launch. Early access was restricted to paid-plan users who opted into Explorer Mode. Replit described a workflow in which the agent formed a hypothesis, searched relevant files, and gathered information before making changes. A live design preview let users see interfaces earlier, before the full application was built.[6]

Agent 3 launched on September 10, 2025, adding code testing and fixing and the creation of custom agents and workflows. Replit described browser-based self-testing in which users could watch the agent check buttons, forms, APIs, and data sources. The company reported operating durations of two minutes for Agent 1, up to 20 minutes for Agent 2, and up to 200 minutes for Agent 3.[7]

Agent 4’s March 2026 release combined visual design iteration with task coordination. Users could generate interface variants on an infinite canvas, edit them visually, and apply a selected design to an app. Independent tasks could run in parallel; larger tasks could be split among sub-agents and recombined, with specialized sub-agents handling conflicts. Users could track tasks, review results, and approve them before merging into the main project. At launch, parallel execution was designated for Pro and Enterprise users, with temporary Core access. Masad said Agent 4 itself had been built using Agent 3.[8][9]

On August 18, 2026, Amjad Masad and Michele Catasta announced Free Mode, identifying OpenAI’s GPT-5.6 Luna as its underlying model. Everyday tasks in that mode did not consume usage credits, but access remained part of Core and Pro subscriptions and subject to limits resetting every five hours, with higher limits for Pro. The announcement also renamed Economy Mode to Power Mode for cost-optimized work and described Max Mode as using higher-performance models for more complex tasks.[11]

Replit announced closed funding of $250 million at a $3 billion valuation on September 10, 2025. Prysm Capital led the round, with Amex Ventures and Google’s AI Futures Fund participating as strategic investors. At the same time, Replit reported annualized revenue rising from $2.8 million to $150 million in less than a year, a community of more than 40 million users, and over 500,000 professional users. Annualized revenue expresses a revenue pace; these user counts do not establish paying-customer totals.[7]

On March 11, 2026, Masad announced that Replit had raised $400 million at a $9 billion valuation. Investors included Georgian, G Squared, Prysm Capital, YC, Coatue, a16z, Craft Ventures, and Qatar Investment Authority, alongside strategic investments from Accenture Ventures, Databricks Ventures, Okta Ventures, and Tether. Masad said the funds would support expansion particularly in Europe, Asia, and the Middle East, future product development, and infrastructure capacity.[9]

In that March announcement, Masad reported more than 50 million users and users from 85% of Fortune 500 companies—an affiliation measure rather than a count of enterprise contracts. He forecast reaching $1 billion in run-rate revenue by the end of 2026; the figure was a target, not an achieved result. Replit also described partnerships supporting deployment and commercialization: Razorpay enabled embedded payments in apps built on Replit, while Hexaware helped enterprises deploy those apps within existing IT systems and workflows.[9]