What are the differences between Leo's AI Models? Follow
Large Language Models (LLMs) are advanced AI systems trained to generate human-like responses. Brave’s AI Assistant, Leo, provides options for users to choose the appropriate model and model configuration for their privacy and functionality needs. These LLMs include Meta's Llama 3 and Mistral AI’s Mixtral models which Brave self-hosts through our own secure infrastructure, as well as models from other companies like Anthropic's Claude 3 Haiku and Claude 3 Sonnet models served via Anthropic’s servers. While the models have different capabilities, Brave ensures user privacy regardless of which model is used and how it is served.
LLMs Explained
Larger models like Llama 3 have more parameters and complexity, allowing them to handle more complex tasks but also requiring more computing power to run efficiently. Models trained on diverse data like Claude 3 Sonnet have broad knowledge, while specialized models like Claude 3 Haiku have domain expertise like coding or legal analysis. Some LLMs excel at text generation, while others are better at question answering or data analysis tasks. Speed will depend on the model size, the power of the infrastructure running the model, the techniques used to optimize it, how much content it is processing, and how complex the job requested is.
Importantly, when Brave provides access to one of our self-hosted LLMs there is no data sharing between Brave and the original model creator. For example, Llama 3 in Brave Leo is managed and operated exclusively by Brave in secure environments, totally independently from the model creator, Meta. Brave also ensures your data is protected when models are served by 3rd parties. Anthropic, for example, can not use Brave Leo user data for training purposes, and user data can not be retained for more than 30 days.
Brave Leo offers options for choosing the appropriate model for your requirements. You can use the Brave-hosted defaults, manually select externally-hosted models from 3rd parties, or, for advanced users, connect Leo to your own model (coming soon…).
Brave-hosted Models
The default models offered to all users in Brave Leo are Brave-hosted. With Brave-hosted models:
- Your IP address is obscured from Brave’s own servers using a reverse proxy. IP addresses are retained only for a short period.
- Your conversations are not stored or logged by Brave after a response is generated. No personal data or conversation transcripts are retained. And your conversations are not used for model training purposes.
- Conversation data is discarded from Brave's servers immediately after the chat session ends. Local conversation data stored on your device can be removed in Settings by selection the option to “Clear Leo data”
Brave-hosted models will change from time to time, typically open-source LLMs running in environments managed using Brave’s secure infrastructure. Currently, Brave-hosted models include the Mixtral model from Mistral AI and Llama 3 from Meta.
Externally-Hosted Models
Some advanced models available in Brave Leo are hosted and served externally by partner companies under contracts that ensure privacy-protection. To access these models, users can manually select them from the Leo settings page.
When using an externally-hosted model:
- We obscure your IP address from the external host using a reverse proxy for added anonymity. IP addresses are retained only for a short period.
- Conversation data is retained by the 3rd party for a limited time as specified in our privacy policy, no more than 30 days. Data is not used to train the AI models.
- Local conversation data stored on your device can be removed in Settings by selection the option to “Clear Leo data”
Externally-hosted models used in Brave products will be served by the LLM providers on their own servers. Currently, Brave provides access to the Claude 3 models by Anthropic, including both Haiku and Sonnet.
Selecting Models
The Leo panel includes a gear icon in the header where you can change the Settings. You will see options for Language Models which you can select or choose to continue with the default.
You can also open advanced settings which will take you to brave://settings/leo-assistant
where you can select which model you want to use by default every time you use Leo.
The models and features available in Brave Leo depend on whether you are using the free version or have a premium subscription:
Feature | Free Leo | Leo Premium |
Rate limits | Reasonable rate limits | Higher rate limits |
Privacy protections |
Inputs are always submitted anonymously through a reverse-proxy and are not retained. Product usage is unlinkable with subscription information. |
Inputs are always submitted anonymously through a reverse-proxy and are not retained. Product usage is unlinkable with subscription information. |
Brave-hosted Models |
Llama 3 by Meta Mixtral by Mistral AI User data is not used to train the AI models. |
Llama 3 by Meta Mixtral by Mistral AI User data is not used to train the AI models. |
3rd Party Models |
Claude 3 Haiku Conversation data is retained for no more than 30 days. User data is not used to train the AI models. |
Claude 3 Haiku Claude 3 Sonnet Conversation data is retained for no more than 30 days. User data is not used to train the AI models. |
Regardless of which model you choose, Brave cannot connect your subscription information to your actual usage of the models, providing an extra layer of privacy for your conversations. The email you used to create your account is unlinkable to your day-to-day use of Leo, making this a uniquely private experience.
Brave's commitment to user privacy extends to its LLM offerings, providing a secure and transparent experience. Whether using Brave-hosted models like Mixtral and Llama 3 or partnering with external providers like Anthropic's Claude models, Brave ensures user data remains private and protected. Users can explore different LLMs with varying capabilities while maintaining control over their personal information.