Codeium

Codeium built AI coding tools that progressed from low-latency autocomplete across multiple editors to Windsurf, an IDE for collaborative, multi-step work with AI. Its consequential choices included offering free code completion, maintaining its own inference and retrieval systems, and supporting enterprise developers in their existing editors. Windsurf launched in November 2024; Cognition announced an agreement to acquire its product and business in July 2025 and introduced its successor, Devin Desktop, in June 2026.[1][2][3][5][7]

The coding products grew out of earlier infrastructure work at Exafunction. That history explains the serving technology Codeium already had when it began training and deploying coding models.[1]

Varun Mohan, Exafunction/Codeium’s CEO, had managed large-scale offline deep-learning infrastructure at Nuro. In his March 2023 account, he said he started Exafunction to make deep-learning workloads easier to serve at scale. Its infrastructure virtualized GPUs and decoupled them from workloads, while compiler optimizations such as operator fusion combined operations to reduce execution time. As transformer workloads became more popular, the team developed infrastructure that it could later use for its own language models.[1]

Mohan recalled that the team’s experience using GitHub Copilot helped establish the value of coding assistance. Engineers writing CUDA and C++ watched Copilot generate difficult template code in a roughly 500-line file, then compiled it successfully. The experience showed them that a coding model could help with rule-heavy work they found difficult, beyond completing simple loops.[1]

Codeium bootstrapped internally with open-source models, gathered permissively licensed public code, and shipped its own model to meet requirements including fill-in-the-middle completion. Serving required a balance between batching requests for GPU utilization and keeping latency low enough for suggestions on each keystroke. Mohan said the team also accepted a tradeoff in training: larger models might have produced better results, but would have increased latency and made fitting the models onto its hardware more difficult.[1]

Efficient serving and heavy optimization made free code completion feasible. Mohan described autocomplete as the first layer of a broader product, with further capabilities, a possible Pro tier, and enterprise sales providing a path beyond it. Giving all software away had not been the original goal. In the interview published March 2, 2023, he estimated probably over 10,000 users, thousands of daily active users, and approximately 4–5% daily growth.[1]

The team built evaluations across multiple languages and codebases, supplementing HumanEval and accounting for the risk of benchmark examples appearing in training data. It used staged A/B rollouts and completion acceptance as product signals. Codeium also adjusted generation length to the developer’s apparent task: Mohan explained that completing an existing if block could be useful, while inventing an unwanted else branch could add mental load. He credited engineering work beyond latency and a blog post co-written with Sean “swyx” Wang in explaining what a usable coding assistant required.[1]

On January 30, 2024, Codeium announced that it had raised a $65 million Series B, led by Kleiner Perkins with Greenoaks and General Catalyst participating. It reported more than 70 supported languages, more than 40 IDE integrations, and over 300,000 individual users. The company said it wrote over 44% of newly committed code for that user base—a code-generation measure rather than a general productivity gain. It named Anduril and Clearwater Analytics as users and Atlassian as a collaborator.[4]

Enterprise offerings included SaaS, VPC, and on-premises deployment, repository context awareness, and optional local model fine-tuning. Codeium’s January 2024 release described SOC2 Type 2 compliance and restrictions against training its public system on customer code. In the later 2024 interview, Mohan explained that individual and enterprise users shared internally maintained indexing and model-serving infrastructure. Cross-editor support remained important: some customers had more than 70% of developers using JetBrains, and Codeium continued serving Eclipse users.[4][2]

Mohan and Anshul Ramachandran learned enterprise sales through direct involvement. Mohan said the team had sold to roughly 30–40 customers before hiring a sales leader; Ramachandran described personally handling hundreds of deal cycles. Deployment engineers then worked alongside account executives on technical deployments and brought customer feedback into product development. Ramachandran reported that the enterprise product reached $10 million in annual recurring revenue in less than a year. He also qualified the retrospective strategy: some choices were deliberate, while others emerged through lost deals, learning, and luck.[2]

Codeium made the Windsurf Editor generally available on November 13, 2024, using a fork of Visual Studio Code. Owning the editor gave the team more control over presentation and access to developer actions. Mohan supplied a concrete example: the team could implement Supercomplete refactor suggestions inside an extension, but lacked an API to display them as intended, so it dynamically generated PNGs. Windsurf provided a surface for the fuller experience while Codeium continued developing its extensions.[3][2]

Cascade combined codebase knowledge, tools, and awareness of developer actions. It could propose changes across files, edit or create files, search with grep, list directories, and execute code. Its proprietary Riptide tool used LLM-based code search. The company called this combination “AI flows”: assistance that could incorporate the developer’s intervening work rather than requiring a fresh explanation each time. Supercomplete applied the same emphasis on intent to predicting edits beyond the next text at the cursor.[3][2]

Mohan said Cascade research had taken approximately nine months to a year. Its evaluations removed changes from historical commits to leave repositories incomplete, withheld commit messages, and tested whether retrieval, planning, and iterative execution could reconstruct code that passed tests. Ramachandran described retrieval evaluations covering roughly 50 results, using files changed together in historical commits to build reference sets. This tested whether the system gathered distributed pieces needed for a change, rather than simply finding one relevant snippet.[2]

Windsurf used a hybrid model strategy. Codeium’s own models handled keystroke-level autocomplete and Supercomplete, while proprietary systems retrieved context and applied changes to codebases. Third-party models supplied high-level planning. Ramachandran revised the earlier preference for first-party models, crediting external providers’ advances with making Cascade and Windsurf possible. Serving the expanded capabilities also increased costs: the launch retained free functionality while reserving the full experience for paid plans over the longer term.[2][3]

In the interview published December 13, 2024, Mohan reported more than 800,000 developers using Codeium’s extensions. The article’s introduction, written after the recording, reported more than one million developers using the company’s products.[2]

On July 14, 2025, Windsurf CEO Jeff Wang said the company had lost its founders and research team the preceding week. Cognition announced a signed definitive agreement to acquire Windsurf’s IP, product, trademark, brand, business, and joining personnel. Scott Wu and Wang’s July 2026 retrospective identified departures to Google and confirmed that the acquisition subsequently closed. They recalled negotiations involving Jeff, Graham, Scott, and Russell over a roughly 72-hour period.[6][5][8]

At the agreement announcement, Cognition reported Windsurf had $82 million ARR, more than 350 enterprise customers, and hundreds of thousands of daily active users, with enterprise ARR doubling quarter over quarter. Wu said he and Wang structured the transaction so all Windsurf employees would participate financially, have vesting cliffs waived, and receive accelerated vesting for their work to date. Wang proposed combining interactive work in Windsurf with delegation to Cognition’s Devin agents.[5][6]

On June 2, 2026, Cognition introduced Devin Desktop as the next generation of Windsurf. It retained a full IDE and added a Kanban interface for managing local and cloud agents. Spaces grouped sessions, pull requests, files, and shared context; Agent Client Protocol support brought in Codex, Claude Agent, OpenCode, and other compatible agents. Cognition’s team rewrote the local agent in Rust as Devin Local, Cascade’s successor.[7]

Wu and Wang’s July 2026 retrospective listed SWE-1.5, SWE-1.6, and SWE-1.7 among the combined team’s model work, alongside Arena Mode and a public evaluation leaderboard. It also reported growth from 44 to 350 people and from $73 million to more than $500 million in revenue run rate since merging the brands. Those figures describe the combined Cognition business, rather than standalone Codeium or Windsurf.[8]