Tabnine

Tabnine is a privacy-focused, enterprise artificial intelligence code assistant that provides inline code completions, chat-based assistance, test generation, and code review inside a developer's integrated development environment (IDE), with deployment options that include fully air-gapped, on-premises installation.^[1] It is widely recognized as one of the earliest commercial products to apply deep learning to code completion: its 2019 "Deep TabNine" release fine-tuned OpenAI's GPT-2 on roughly two million GitHub files, predating GitHub Copilot by about two years.^[4][5] As of 2025, Tabnine reports serving more than one million developers and generating over one percent of the world's code.^[1][12]

The company traces its corporate roots to Codota, founded in Tel Aviv, Israel, in 2013 by Dror Weiss and Eran Yahav.^[7] Codota acquired the independently built TabNine code completion tool in December 2019 and rebranded the entire company to Tabnine in May 2021.^[3][10] Tabnine differentiates itself from competitors primarily through its focus on privacy, security, and enterprise control: it offers SaaS, virtual private cloud (VPC), on-premises Kubernetes, and fully air-gapped deployments, and its proprietary "Protected" models are trained exclusively on permissively licensed open-source code to eliminate the legal risk that suggestions could match proprietary or copyleft codebases.^[21][23]

History

Codota (2013 to 2019)

Codota was founded in 2013 by Dror Weiss and Professor Eran Yahav in Tel Aviv, Israel.^[7] Weiss, a software engineer with a background at Panaya (an enterprise SaaS company), served as CEO. Yahav, an associate professor of Computer Science at the Technion (Israel Institute of Technology), served as CTO.^[25] The company grew out of over a decade of academic research at the Technion into program synthesis, program analysis, and machine learning for code.^[25]

Codota's original product used a semantic approach to code assistance. Rather than treating code as raw text, it built models that captured the meaning and structure of code patterns. The tool analyzed millions of open-source code repositories to learn common programming patterns and offer context-aware suggestions to developers.^[8] Initially, Codota focused on Java development and was available as a plugin for JetBrains IDEs.^[8]

In June 2017, Codota raised $2 million in a seed round from Khosla Ventures and Bob Pasker's Syndicate. This was followed by a $12 million Series A round in April 2020, led by e.ventures with participation from Khosla Ventures.^[7] At the time of the Series A, Codota reported that its user base had grown over 1,000 percent in the prior year, reaching more than one million monthly developers.^[8]

Who created the original TabNine? (2018 to 2019)

The TabNine product itself was created independently by Jacob Jackson, a computer science undergraduate student at the University of Waterloo in Canada.^[28] Jackson, a former intern at both Jane Street and OpenAI, started building TabNine in February 2018 while working at Jane Street.^[5] He released the first version in November 2018 as a code completion plugin.^[5]

In mid-2019, Jackson released "Deep TabNine," which integrated OpenAI's GPT-2 model.^[4] This was one of the first commercial applications of a large language model to code generation. Deep TabNine used GPT-2 (a transformer-based natural language processing model with 1.5 billion parameters) fine-tuned on approximately two million files from GitHub to predict code tokens.^[6] The tool treated code as text and predicted each token given the tokens that preceded it, which differed from Codota's semantic approach.^[6] Jackson described the underlying mechanism plainly: "You feed it a sequence of tokens. You can think of one token as one word, and if you have a sequence of words, then it will give you a distribution of the words you're going to see next."^[5] He said his aim was a tool "in between, where it was really fast but smarter than the other tools out there."^[5]

Deep TabNine attracted significant attention in the developer community for its ability to generate multi-line code completions across dozens of programming languages.^[6] It was one of the earliest demonstrations that large-scale language models could write functional code.^[4]

Acquisition and Merger (2019 to 2021)

On December 16, 2019, Codota acquired TabNine.^[10] The terms of the deal were not publicly disclosed.^[9] The acquisition brought together two complementary technical approaches: Codota's semantic code understanding and TabNine's textual, GPT-2-based code prediction. According to CEO Dror Weiss, "a thoughtful combination of the two is far superior to any of the models individually."^[9]

Following the acquisition, Jacob Jackson joined Codota as an advisor and provided training to the engineering team. He later departed and went on to found Supermaven, another code completion startup, which was subsequently acquired by Anysphere (the company behind Cursor).^[26]

For roughly 18 months after the acquisition, Codota operated both the Codota and TabNine products in parallel. As the team integrated Codota's Java-focused capabilities into the TabNine platform, the Codota product became redundant.^[2]

On May 26, 2021, Codota officially rebranded to Tabnine, adopting the name of its more widely recognized product.^[3] The rebranding coincided with the release of the company's first proprietary large language model for code.^[3] Existing Codota users, particularly those on JetBrains and Eclipse, were migrated to the Tabnine plugin.^[3] By April 2022, Tabnine reported surpassing one million users.^[1]

Growth and Enterprise Focus (2022 to Present)

Throughout 2022 and 2023, Tabnine shifted its strategic focus toward enterprise customers. The company introduced its Enterprise tier with advanced features including on-premises deployment, codebase personalization, and administrative controls for security-conscious organizations.^[11]

In June 2022, Tabnine raised $15.5 million in a funding round co-led by Qualcomm Ventures, OurCrowd, and Samsung NEXT.^[27] In November 2023, the company closed a $25 million Series B round led by Telstra Ventures, with participation from Atlassian Ventures, Elaia, Headline, Hetz Ventures, Khosla Ventures, and TPY Capital.^[11][12]

In 2024, Tabnine undertook a restructuring that reduced its workforce by approximately 18 percent (15 employees from a team of about 80), primarily affecting marketing roles in Israel and the United States.^[18] The company framed the layoffs as a strategic shift to prioritize resources for enterprise growth, which had been showing strong momentum.^[18]

In July 2024, Tabnine released Tabnine Protected 2, a second-generation proprietary model trained exclusively on permissively licensed code.^[13] The company reported that Protected 2 delivered performance on par with the larger GPT-3.5 Turbo on internal evaluations using the HumanEval and MultiPL-E benchmarks, and supported over 600 programming languages and frameworks, up from approximately 80 in the previous version.^[13][14]

In September 2025, Tabnine was named a Visionary in the Gartner Magic Quadrant for AI Code Assistants, an improvement from its Niche Player position in 2024.^[16] In November 2025, Tabnine launched its Enterprise Context Engine and Org-Native Agents platform, marking a shift toward agentic AI capabilities.^[15] The company also won the InfoWorld Technology of the Year Award 2025 in the Software Development Tools category for the second time (having previously won in 2023).^[17]

Founders

Scroll sideways for more →

How does Tabnine's technology work?

Early Architecture: GPT-2

When Jacob Jackson created Deep TabNine in 2019, it was built on top of OpenAI's GPT-2 model.^[4] GPT-2, a transformer-based language model released by OpenAI in February 2019, was originally designed for natural language text generation. Jackson fine-tuned GPT-2 on a large corpus of source code from GitHub, training the model to predict code tokens in sequence.^[6] This approach treated code as a stream of text tokens rather than analyzing its semantic structure.^[6] Jackson noted that the main obstacle was performance: "The biggest challenge in applying deep learning is that these models are computationally intensive, and because we require a high-performing and highly responsive system, that's a problem."^[5]

Deep TabNine was notable for being one of the very first products to apply a large-scale transformer model to code completion. It preceded GitHub Copilot (launched June 2021) and Amazon CodeWhisperer (launched June 2022) by roughly two to three years.^[4]

Proprietary Models

After the Codota acquisition, Tabnine gradually moved away from using OpenAI's models and developed its own proprietary language models for code.^[21] As of 2025, Tabnine offers two categories of proprietary models:

Tabnine Protected (Universal) Models: These are Tabnine's proprietary models trained exclusively on open-source code with permissive licenses.^[21] The permitted license types include:

Scroll sideways for more →

By restricting training data to these licenses, Tabnine ensures that code suggestions cannot match proprietary or copyleft-licensed code, reducing legal liability for enterprise users.^[21] The latest version, Tabnine Protected 2 (released July 2024), supports over 600 programming languages and frameworks and was benchmarked on HumanEval and MultiPL-E, where Tabnine reported parity with GPT-3.5 Turbo.^[13][14]

Third-Party Model Integration: In addition to its proprietary models, Tabnine allows users (particularly on Enterprise plans) to connect to third-party models from providers including:^[20]

Scroll sideways for more →

Enterprise administrators can control which models are available to their organization.^[20] This model list changes frequently as new models are released.

Enterprise Context Engine

Launched in November 2025, the Enterprise Context Engine is a proprietary system that continuously analyzes and models an organization's software environment.^[15] It combines vector retrieval, graph-based analysis, and agentic retrieval techniques to interpret relationships across codebases, tools, and project management tickets.^[15] This context layer feeds into Tabnine's AI features (completions, chat, and agents), allowing the tool to produce suggestions that are relevant to the specific architecture, coding conventions, and dependencies of each organization.^[15]

The Enterprise Context Engine can connect to an organization's repositories, services, APIs, documentation, and CI/CD systems. Unlike static training, it adapts to new codebases and policies without requiring retraining or redeployment of the underlying models.^[15]

What can Tabnine do?

Code Completion

Tabnine's core feature is inline code completion. As a developer types, the IDE plugin sends context from the active file and connected repositories to Tabnine's model, which returns suggestions ranging from single-line completions to entire functions, tests, and docstrings.^[1] Tabnine reports that it typically automates 30 to 50 percent of code generation tasks for active users.^[1][12]

Code completion works across all supported IDEs and supports over 600 programming languages.^[19] The tool adapts to the coding patterns and conventions found in the developer's codebase, providing personalized suggestions over time.

Chat

Tabnine Chat provides a conversational interface within the IDE where developers can ask questions in natural language. Capabilities include:

• Generating code from natural language descriptions
• Explaining existing functions and code blocks
• Suggesting refactoring approaches
• Creating unit tests for selected code
• Answering questions about the codebase

Chat is available in VS Code, JetBrains IDEs, Visual Studio, and Eclipse.^[19]

AI Agents

In 2025, Tabnine introduced Org-Native Agents that can plan, execute, and validate multi-step development tasks, positioning the product as an AI coding agent platform.^[15] These agents leverage the Enterprise Context Engine to understand organizational context.^[15] Available agents include:

Scroll sideways for more →

The Code Review Agent won "Best Innovation in AI Coding" at the 2025 AI TechAwards for its ability to catch defects at the pull request level.

Code Privacy and Zero Data Retention

A central part of Tabnine's value proposition is its privacy model. The company commits to:^[22]

• Never storing customer code (zero data retention)
• Never sharing customer code or usage data with third parties
• Never using customer code to train public models
• Holding code passed to servers only in memory, deleting it once a response is returned

These guarantees apply across all deployment modes.^[22] Enterprise customers on self-hosted deployments have the additional assurance that no code ever leaves their network.^[22]

Supported IDEs

Tabnine provides official plugins for the following IDEs:^[19]

Scroll sideways for more →

The JetBrains plugin covers the full family of IntelliJ Platform IDEs:

Scroll sideways for more →

Plugins for other editors (such as Vim, Neovim, Emacs, and Sublime Text) exist as legacy or community-maintained versions but do not support advanced completions or Tabnine Chat.^[19]

Supported Programming Languages

Tabnine Protected 2 supports over 600 programming languages and frameworks.^[13] The strongest support is available for the most popular languages:

Scroll sideways for more →

Additional languages and frameworks receive varying levels of support. When using third-party models (such as Claude or GPT-5), language support extends to whatever languages those models handle.

What deployment options does Tabnine offer?

Tabnine offers four deployment modes, which is a key differentiator from most competitors that operate exclusively as cloud services:^[23]

Scroll sideways for more →

For on-premises and air-gapped deployments, Tabnine provides Helm charts for installation on Kubernetes clusters.^[23] In air-gapped environments, Docker images are provisioned manually through an internal registry or side-loaded onto the Kubernetes host.^[23] Tabnine's documentation recommends a single NVIDIA GPU (such as an NVIDIA L40S or A100 80GB) to serve up to approximately 1,000 users, with two to four GPUs recommended for larger teams.^[23]

How much does Tabnine cost?

As of 2025, Tabnine offers three pricing tiers:^[24]

Scroll sideways for more →

Tabnine discontinued its free Basic tier in April 2025, replacing it with the more limited Dev Preview plan.^[24] Enterprise pricing may vary based on deployment model, support tier, and team size, and is typically negotiated through a sales process.

Funding

Tabnine has raised capital across multiple rounds since its founding as Codota:^[27]

Scroll sideways for more →

Total reported funding ranges from approximately $57 million to $102 million depending on the source and methodology used.^[27] The discrepancy arises from undisclosed round amounts and varying classifications of funding types across databases such as Crunchbase, PitchBook, and Tracxn.^[27]

How does Tabnine differ from GitHub Copilot and Cursor?

Tabnine operates in an increasingly competitive market for AI coding assistants. Its primary competitors include:

Scroll sideways for more →

Tabnine's competitive positioning centers on enterprise control and privacy. Unlike GitHub Copilot and most other competitors, Tabnine offers fully on-premises and air-gapped deployment, which appeals to organizations in regulated industries (finance, healthcare, defense, government) that cannot send code to third-party cloud services.^[23] Additionally, Tabnine's Protected models, trained only on permissively licensed code, address intellectual property concerns that some organizations have with tools trained on broader datasets.^[21] Whereas Cursor competes on an AI-native editor experience aimed largely at individual developers, Tabnine targets the security and compliance needs of large engineering organizations.

However, Tabnine faces challenges from the scale and resources of larger competitors. GitHub Copilot, backed by Microsoft, surpassed 20 million total users by mid-2025. Cursor has gained rapid popularity among individual developers for its AI-native editor experience. Amazon Q Developer benefits from deep integration with the AWS ecosystem. The competitive landscape continues to evolve as new entrants and approaches (such as autonomous coding agents) emerge.

Industry Recognition

Tabnine has received several industry awards and analyst recognition:

Scroll sideways for more →

Limitations

• Tabnine's proprietary Protected models, while safe from licensing concerns, may produce less varied suggestions than models trained on broader datasets.
• On-premises and air-gapped deployments require dedicated GPU hardware (NVIDIA GPUs recommended), adding infrastructure costs.^[23]
• IDE support for advanced features (chat, agents) is limited to VS Code, JetBrains, Eclipse, and Visual Studio. Users of other editors receive only legacy code completion.^[19]
• The free Dev Preview tier is more limited than competitors' free offerings (such as Codeium's free unlimited completions).^[24]
• Tabnine's team is significantly smaller (approximately 70 to 80 employees) than the engineering teams behind competing products at Microsoft, Amazon, and Google.^[18]

See Also

• GitHub Copilot
• Cursor
• Codeium
• Amazon Q Developer
• AI Coding Agent
• Large Language Model
• GPT-2
• Transformer Architecture
• AI Code Generation

References

1. Tabnine. "Tabnine AI Code Assistant." https://www.tabnine.com/ ↩
2. Tabnine Blog. "Tabnine is now part of Codota." March 23, 2020. https://www.tabnine.com/blog/tabnine-part-of-codota/ ↩
3. Tabnine Blog. "Codota is now Tabnine!" May 26, 2021. https://www.tabnine.com/blog/codota-is-now-tabnine/ ↩
4. Synced Review. "Deep TabNine: A Powerful AI Code Autocompleter For Developers." July 18, 2019. https://syncedreview.com/2019/07/18/deep-tabnine-a-powerful-ai-code-autocompleter-for-developers/ ↩
5. IEEE Spectrum. "Q&A: This Autocompletion Tool Aims to Supercharge Your Coding." https://spectrum.ieee.org/qa-this-autocompletion-tool-aims-to-supercharge-your-coding ↩
6. The Register. "Try out this deep-learning AI bot that autocompletes lines of source code for you." July 22, 2019. https://www.theregister.com/2019/07/22/ai_coding_bot/ ↩
7. CTech (Calcalist). "Israeli Startup Codota Raises $12 Million in Series A Round Led By Khosla Ventures." April 2020. https://www.calcalistech.com/ctech/articles/0,7340,L-3812190,00.html ↩
8. VentureBeat. "Codota raises $12 million for AI that suggests and autocompletes code." April 2020. https://venturebeat.com/ai/codota-raises-12-million-for-ai-that-suggests-and-autocompletes-code/ ↩
9. Globes. "Israeli AI code developer Codota buys Canada's TabNine." December 2019. https://en.globes.co.il/en/article-israeli-startup-codota-buys-canadas-tabnine-1001311136 ↩
10. SiliconANGLE. "Israeli AI-assisted software development firm Codota acquires TabNine." December 16, 2019. https://siliconangle.com/2019/12/16/israeli-ai-assisted-software-development-firm-codota-acquires-tabnine/ ↩
11. TechCrunch. "Code-generating AI platform Tabnine nabs $25M investment." November 8, 2023. https://techcrunch.com/2023/11/08/code-generating-ai-platform-tabnine-nabs-25m-investment/ ↩
12. Tabnine Blog. "Tabnine raises $25M Series B funding." November 2023. https://www.tabnine.com/blog/tabnine-series-b/ ↩
13. GlobeNewsWire. "Tabnine Unveils Second Generation Protected LLM." July 25, 2024. https://www.globenewswire.com/news-release/2024/07/25/2918855/0/en/Tabnine-Unveils-Second-Generation-Protected-LLM-to-Keep-AI-Workloads-Private-Protected-and-Compliant.html ↩
14. Tabnine Blog. "Announcing Tabnine Protected 2." https://www.tabnine.com/blog/announcing-tabnine-protected-2-a-license-safe-llm-that-performs-as-strong-as-the-best/ ↩
15. GlobeNewsWire. "Tabnine Launches Enterprise-Fit Agentic AI Powered by Its Enterprise Context Engine." November 5, 2025. https://www.globenewswire.com/news-release/2025/11/05/3181534/0/en/Tabnine-Launches-Enterprise-Fit-Agentic-AI-Powered-by-Its-Enterprise-Context-Engine.html ↩
16. GlobeNewsWire. "Tabnine Named a Visionary in the 2025 Gartner Magic Quadrant for AI Code Assistants." September 17, 2025. https://www.globenewswire.com/news-release/2025/09/17/3151877/0/en/Tabnine-Named-a-Visionary-in-the-2025-Gartner-Magic-Quadrant-for-AI-Code-Assistants.html ↩
17. GlobeNewsWire. "Tabnine Wins InfoWorld Technology of the Year Award 2025." January 6, 2026. https://www.globenewswire.com/news-release/2026/01/06/3213708/0/en/Tabnine-Wins-InfoWorld-Technology-of-the-Year-Award-2025-for-Software-Development-Tools.html ↩
18. CTech (Calcalist). "Tabnine cuts 18% of workforce in restructuring effort." 2024. https://www.calcalistech.com/ctechnews/article/h1jlmiek1l ↩
19. Tabnine Docs. "Supported IDEs." https://docs.tabnine.com/main/welcome/readme/supported-ides ↩
20. Tabnine Docs. "AI Models." https://docs.tabnine.com/main/welcome/readme/ai-models ↩
21. Tabnine Docs. "Tabnine's Private and Protected Universal Models." https://docs.tabnine.com/main/welcome/readme/ai-models/tabnines-private-and-protected-universal-models ↩
22. Tabnine. "Total AI code privacy & zero data retention." https://www.tabnine.com/code-privacy/ ↩
23. Tabnine Docs. "Deployment Options." https://docs.tabnine.com/main/welcome/readme/architecture/deployment-options ↩
24. Tabnine. "Plans & Pricing." https://www.tabnine.com/pricing/ ↩
25. Technion Computer Science. "Eran Yahav." https://csaws.cs.technion.ac.il/~yahave/ ↩
26. TechCrunch. "AI coding assistant Supermaven raises cash from OpenAI and Perplexity co-founders." September 16, 2024. https://techcrunch.com/2024/09/16/ai-coding-assistant-supermaven-raises-cash-from-openai-and-perplexity-founders/ ↩
27. Clay. "How Much Did Tabnine Raise? Funding & Key Investors." https://www.clay.com/dossier/tabnine-funding ↩
28. BetaKit. "Waterloo startup TabNine acquired by Israeli startup Codota." https://betakit.com/waterloo-startup-tabnine-acquired-by-israeli-startup-codota/ ↩