CtrlK
Supermaven was an AI code completion tool that used a proprietary neural network architecture to deliver fast, context-aware code suggestions directly inside a developer's editor. Founded in 2024 by Jacob Jackson, the creator of Tabnine, Supermaven distinguished itself from competitors through its unusually large context window (initially 300,000 tokens, later expanded to one million tokens) and its low-latency inference, which the company reported was roughly three times faster than competing tools such as GitHub Copilot, Codeium, and Tabnine. Supermaven was based in New York and operated as a standalone product until November 2024, when it was acquired by Anysphere, the company behind the AI code editor Cursor.
At its peak, Supermaven reported over 20,000 active developers using the tool. The product offered both a free tier and a paid Pro plan at $10 per month, with extensions available for Visual Studio Code, JetBrains IDEs, Neovim, and Zed.
Jacob Jackson, a computer science student at the University of Waterloo in Canada, began building the original TabNine in February 2018 during an internship at Jane Street, a quantitative trading firm. He released the first version in late 2018 as a code completion plugin. In mid-2019, Jackson added deep learning completions powered by GPT-2, creating one of the earliest commercial applications of a large language model to code generation. TabNine was acquired by Codota (later renamed Tabnine) in December 2019.
After selling TabNine, Jackson joined OpenAI as a research scientist, where he worked on semi-supervised learning and distributed training for large language models for approximately two and a half years. He left OpenAI around 2022.
After leaving OpenAI, Jackson began working on a new code completion tool. He officially launched Supermaven on February 22, 2024, announcing it as "the first code completion tool with a 300,000-token context window." The launch was accompanied by a post on Hacker News and drew immediate attention from the developer community for its speed claims and the large context window, which was roughly 37 times larger than GitHub Copilot's 8,192-token context window at the time.
Jackson argued that building custom models and serving infrastructure from scratch gave Supermaven a structural cost advantage. He noted that GPU serving costs, not just model development expenses, were the main bottleneck for code completion tools, and that an efficient small company could compete against larger players like Microsoft and Google by optimizing the full inference stack.
On September 16, 2024, Supermaven announced a $12 million Series A financing round led by Bessemer Venture Partners. The round included angel investments from several prominent figures in the AI industry:
| Investor | Affiliation |
|---|---|
| Bessemer Venture Partners | Lead investor |
| John Schulman | Co-founder, OpenAI |
| Denis Yarats | Co-founder, Perplexity AI |
| Eoghan McCabe | Founder, Intercom |
| Mike Volpi | Co-founder, Index Ventures San Francisco |
The company said it planned to use the funds to develop a new text editor designed to help experienced developers navigate and edit large codebases, as well as to expand its model capabilities.
On July 2, 2024, Supermaven announced version 1.0, which included the launch of Babble, a new proprietary model that was 2.5 times larger than the previous Supermaven model. Babble expanded the context window from 300,000 tokens to one million tokens and was trained on a larger corpus of code. Despite its increased size, Babble achieved lower latency than its predecessor thanks to simultaneous improvements in Supermaven's serving infrastructure.
The 1.0 release also introduced in-editor chat functionality supporting GPT-4o and Claude 3.5 Sonnet, along with a team billing system for organizations.
On November 12, 2024, Anysphere announced that it had acquired Supermaven. The acquisition price was not publicly disclosed. The deal came at a time when Anysphere was receiving unsolicited investment offers valuing the company at as much as $2.5 billion from firms including Benchmark, Index Ventures, and Andreessen Horowitz.
In a blog post announcing the acquisition, Jackson explained that the Supermaven team had realized the next step for their product required more than better AI models. They needed to co-design models with the user interface, which meant building their own editor, because the extension APIs in VS Code and JetBrains IDEs were "too restrictive to build the interface we wanted." After discussions with Cursor's team, both sides concluded that joining forces would produce a more useful product than either company could build independently.
Anysphere CEO Michael Truell stated that Supermaven's technology would enable a new version of Cursor's Tab AI model that would be "fast, context-aware, and highly intelligent," with particular strength in handling long sequences of code.
Following the acquisition, Supermaven's existing plugins for VS Code, JetBrains, and Neovim continued to be maintained, though the team's primary focus shifted to developing Cursor.
Supermaven used a proprietary neural network architecture rather than a standard Transformer. Jackson described the architecture as being more efficient than a Transformer for processing long code contexts. Specifically, the architecture allowed Supermaven to offer a 300,000-token context window (later one million tokens) while keeping the computational cost and latency comparable to a Transformer with a 4,000-token context window.
The key design principle was that code completion required a different approach than general-purpose language modeling. By building both the model and the serving infrastructure from scratch, Supermaven could optimize the entire pipeline end-to-end for the specific task of inline code completion.
Babble was Supermaven's proprietary large language model for code completion. The model went through at least two major versions:
| Version | Context Window | Release | Notes |
|---|---|---|---|
| Original model | 300,000 tokens | February 2024 | Launched with Supermaven's initial release |
| Babble (1.0) | 1,000,000 tokens | July 2024 | 2.5x larger than predecessor; trained on larger code corpus |
Babble's architecture was designed to handle long-range dependencies in code. The model could process an entire repository's worth of code and recall specific details regardless of their position within the context. In a "needle in a haystack" benchmark, Babble demonstrated 100% recall of information hidden within one million tokens of context.
Supermaven did not publicly disclose the exact parameter count or full architectural details of Babble, describing the architecture only as a "new neural architecture" that was distinct from Transformers.
Unlike most code completion models that train on static code files, Supermaven's model was trained on sequences of code edits, similar to the changes captured by git diff. This approach taught the model how developers actually modify code over time rather than treating each file as an isolated snapshot. Jackson argued that this training methodology produced more natural and contextually relevant suggestions because it reflected real development workflows.
Supermaven was the first code completion tool to offer "next location prediction," a feature where the model predicted not only what code to write but also where to write it. This meant Supermaven could suggest jumping the cursor to a different location in the file and inserting code there, which reduced the number of manual navigation steps a developer needed to take.
Supermaven's serving infrastructure was optimized for low-latency responses. The model was triggered on nearly every keystroke, and the system was designed to return suggestions within approximately 100 milliseconds. In a latency comparison published on Supermaven's blog at launch, the tool showed the following results:
| Tool | Average Latency | Frames (at 60fps) |
|---|---|---|
| Supermaven | 250ms | 15 |
| GitHub Copilot | 783ms | 47 |
| Tabnine | 833ms | 50 |
| Codeium | 883ms | 53 |
| Cursor (pre-acquisition) | 1,883ms | 113 |
These benchmarks were self-reported by Supermaven and measured the time between a keystroke and the appearance of a suggestion in the editor.
Supermaven published three internal benchmarks to evaluate the effectiveness of its large context window:
Needle in a Haystack: A random key-value pair was inserted at various positions within 50,000 to 300,000 tokens of code. The model was tested on its ability to retrieve the correct value after processing the full context. Testing involved 20 runs per token count with the needle placed at different positions each time. The model achieved strong accuracy across all tested context lengths.
Dense Retrieval: Repeated key-value pairs were placed throughout a 300,000-token sequence. The model needed to remember all information to complete assertions on second occurrences. Results showed near-100% accuracy when pairs were separated by fewer than 10,000 tokens or more than 290,000 tokens, with lowest accuracy (around 75%) when occurrences were separated by approximately 50,000 tokens.
Prediction Error vs. Context Length: Using 240,000 tokens of out-of-sample internal code, prediction accuracy was measured as a function of token position across 545 random file orderings. Error decreased smoothly as context increased, confirming that the model made use of its full context window.
Supermaven's primary feature was inline code completion. As a developer typed, the extension sent context from the active file and the broader repository to Supermaven's model, which returned suggestions in real time. After spending 10 to 20 seconds processing a repository at startup, the tool became familiar with the developer's APIs and coding conventions.
Supermaven could generate both single-line and multi-line completions. The model's large context window meant it could reference files and definitions across the entire project when generating suggestions, rather than being limited to the currently open file.
On June 25, 2024, Supermaven launched a chat feature within its VS Code and JetBrains extensions. Chat allowed developers to interact with AI models (GPT-4o and Claude 3.5 Sonnet) directly in their editor without switching to a browser or separate application. Key chat capabilities included:
Chat was included at no additional cost for Pro subscribers. Free tier users could access chat by providing their own OpenAI or Anthropic API keys.
Supermaven Pro included style adaptation, where the model learned from a developer's coding patterns over time and adjusted its suggestions to match their preferred conventions and style. This feature used the extended context window to identify recurring patterns in how a developer wrote code.
Supermaven provided official extensions for the following development environments:
| Editor | Extension Type | Chat Support |
|---|---|---|
| Visual Studio Code | VS Code Extension (v0.2.10+) | Yes |
| JetBrains IDEs | Plugin (v1.30+) | Yes |
| Neovim | Plugin | Completion only |
| Zed | Built-in integration | Completion only |
The JetBrains plugin supported the full family of IntelliJ-based IDEs, including IntelliJ IDEA, PyCharm, WebStorm, GoLand, CLion, PhpStorm, Rider, and RubyMine.
Supermaven offered two individual plans and a team plan:
| Plan | Price | Context Window | Chat Credits | Style Adaptation | Data Retention |
|---|---|---|---|---|---|
| Free | $0/month | Standard | None (bring your own API key) | No | 7 days |
| Pro | $10/month | 1,000,000 tokens | $5/month included | Yes | 7 days |
| Team | $10/user/month | 1,000,000 tokens | $5/month per user | Yes | 7 days |
The Pro plan included a 30-day free trial. The Team plan added centralized user management and billing for organizations.
Supermaven entered a competitive market for AI coding assistants. Its main differentiators were speed and context window size. The following table compares Supermaven with its primary competitors at the time of its operation:
| Tool | Developer | Context Window | Price (Individual) | Key Differentiator |
|---|---|---|---|---|
| Supermaven | Supermaven (later Anysphere) | 1,000,000 tokens | Free / $10/month | Fastest latency; largest context window |
| GitHub Copilot | GitHub / Microsoft | 8,192 tokens (at launch) | Free / $10-$39/month | Largest user base; deep GitHub integration |
| Tabnine | Tabnine | Varies | $9/month | Privacy-focused; on-premises deployment |
| Codeium | Exafunction | Varies | Free / $15/month | Free unlimited completions |
| Cursor | Anysphere | Varies | $20/month | AI-native code editor; strong agentic features |
| Amazon Q Developer | Amazon Web Services | Varies | Free / $19/month | AWS integration |
Although Supermaven operated independently for less than a year (February to November 2024), it had a notable impact on the AI code completion market. Its emphasis on large context windows pushed competitors to expand their own context capabilities. The tool's latency benchmarks established a new standard for how fast code suggestions should appear.
Jacob Jackson's career trajectory, from creating Tabnine as an undergraduate in 2018 to building Supermaven and joining Cursor, places him among the most influential individual contributors to the AI code completion field. His work at each stage (TabNine, OpenAI, Supermaven) introduced new ideas that were later adopted across the industry: applying GPT-2 to code, training on edit sequences rather than static files, and building custom neural architectures optimized specifically for code completion latency.
After the acquisition, Supermaven's technology was integrated into Cursor's Tab model. The Supermaven extensions for VS Code, JetBrains, and Neovim continued to receive updates but were no longer the team's primary development focus.