AI models
Tabnine's AI coding assistance is backed by Tabnine’s proprietary AI models for code completions and chat, which are trained and hosted by Tabnine and are private and protected.
In addition, Tabnine Chat includes the option of using third-party models. The privacy policies and the protection offered by these third-party models may be different from the Tabnine models.
Tabnine’s AI models
These are the proprietary models hosted by Tabnine:
Tabnine Universal code completions model: Tabnine’s proprietary model is designed to deliver exceptional performance without the risk of intellectual property violations. It's trained and hosted by Tabnine and is available in all tiers.
Tabnine Protected chat model: Tabnine’s core model is designed to deliver high performance without the risk of intellectual property violations.
Optional AI models for chat
Tabnine Chat users can choose from these chat models (in addition to the Tabnine Protected chat model):
Tabnine + Mistral: Tabnine’s newest offering is built to deliver the highest possible performance while maintaining complete privacy. It's hosted by Tabnine.
GPT-4o, GPT-4 Turbo, and GPT-3.5 Turbo: The industry’s most popular LLMs, proven to deliver the highest levels of performance for teams willing to share their data externally.
Claude 3.5 Sonnet: This model raises the bar for intelligence and outperforms its predecessor, Claude 3 Opus, on a wide range of evaluations, including coding. It operates at twice the speed of Claude 3 Opus. This model is recommended for users looking for the best possible performance who are comfortable sharing their data externally and using models trained on any publicly available code.
Codestral: Trained on more than 80 programming languages, Codestral — Mistral’s first-ever code model, demonstrates proficiency in both widely used and less-common languages. This model is recommended for users looking for the best possible performance who are comfortable sharing their data externally and using models trained on any publicly available code.
Command R: Cohere Command R model is ideal for large-scale production workloads and balances high efficiency with strong accuracy. This model is recommended for users looking for the best possible performance who are comfortable sharing their data externally and using models trained on any publicly available data.
Tabnine gives you the insight you need to choose
Tabnine users can choose which chat model to use. This decision depends on the specific use case and constraints of each user around these three main aspects:
Performance: Does the model provide accurate, relevant results for the programming languages and frameworks I’m working in right now?
Privacy: Does the model store my code or user data? Could my code or data be shared with third parties? Is my code used to train their model?
Protection: What code was the model trained on? Is it all legally licensed from the author? Will I create risks for my business by accepting generated code from a model trained on unlicensed repositories?
This comparison table will help you to decide which model is right for you:
Performance levels
The performance levels below are Tabnine’s estimation of how each model behaves in real-world software development use cases, as Tabnine has deployed them with context awareness.
Good = This class of models performs acceptably for the majority of software development use cases. They may not have comprehensive support for a large number of languages and frameworks and may suffer from a significant number of hallucinations.
Better = This class of models performs to a high standard that allows for consistently safe use for most software development use cases. They may struggle with obscure languages and frameworks, and a limited amount of hallucinations may still be observed.
Best = This class of models performs to the highest standard observed among the currently available models. They offer the most comprehensive support for the largest variety of languages and frameworks, and a limited amount of hallucinations may still be observed.
Privacy
Private = No code data is retained or shared with Tabnine or any other entities.
Not private = Code or data may be shared with third parties, as per their public terms of service. Tabnine still adheres to our zero data retention policy.
Protection
Protected = The model was exclusively trained on code with permissive open source licenses, or on code that was otherwise licensed by the model provider. Any code used for training is explicitly allowed for use by developers without encumbrances.
Not protected = Model training data may include code with licenses that do not explicitly allow their reuse or allow their use for training AI models.
Tabnine users can choose which Tabnine Chat model to use
Tabnine Pro users specify their preferred model the first time they use Chat, and can change it at any time. For projects where data privacy and legal risks are less important, you can use a model optimized for performance over compliance. As you switch to working on projects that have stricter requirements for privacy and protection, you can change to a model like Tabnine Protected that's built for that purpose. The underlying LLM can be changed with just a few clicks — and Tabnine Chat adapts instantly.
Tabnine Enterprise administrators control and specify the models that are available to their organization. Administrators control the available models for their organization. Enterprises often make strategic bets on using specific models across their organization. This update helps Tabnine to be compatible with your chosen LLM and be a part of its ecosystem and makes it easier for you to get the most out of Tabnine without evolving your LLM strategy.
Selecting and switching between chat models
Fine-tuned models for Enterprise customers
Enterprise customers have the option to deploy private fine-tuned models. Fine-tuned models are private models that are the result of refining the Universal completion model or the Tabnine Protected chat model with the customer codebase. For code completions, the fine-tuned model replaces the Universal model, and for Chat it’s added as another private and protected option for the model.
Last updated
