Hi. Thank you for the new feature Ask Brave.
Could you clarify which AI model is used behind the scenes? While the information may be confidential, it would be helpful for Brave users, who prioritize privacy, to know whether it’s self-hosted or proxied.
Additionally, what are the recommended use cases for brave://leo-ai vs https://search.brave.com/ask? The two products seem similar, so it’s not clear in what scenarios each chatbot would be more appropriate. Leo is useful for summarizing the current web page and discussing its content. Ask has the deep research feature and can provide answers based on search results. Is there anything else?
@diannfuchsia good question.
The Ask Brave feature is specific to the Search engine. It is using a modified version of Brave Leo, built specifically for Search. As for your privacy, all chats are encrypted, ephemeral and expire by default after 24 hours of inactivity. We do not retain IP addresses. More information on this here:
As for which one to use in which scenario – Leo AI built into the browser is best use when you want information about the site you’re viewing, or if you want to have a long-form chat/conversation with the AI. Ask Brave is tailored to Brave Search results that are most relevant to the original query asked, similar to Googles “AI mode” in Google Search.
As of today it’s using a combination of llama and qwen, always open models we can run in-house to guarantee privacy. Most of the work is not on the model but on the selection of what data is chosen as context. We switch models very, very often So this will be outdated soon.
That was back in January. I imagine the models have changed up. But as they said, they run things for maximum privacy.
The latest iteration of the AI, now called “Ask Brave” is inferior in tone if not in accuracy to the old one. It’s more likely to have what I’d call a snarky tone. I’ve frequently found it is more likely to simply fetch information I’d characterize as “conventional narrative” even given prompts which should guide it away from such. It’s more prone to slip into what I call “fetch mode”, and simply regurgitate searches, failing to contextualize the previous and subsequent queries during iterative investigation of ideas.
Tonally, it is more aloof. The old version felt like interacting with a trusted study partner, and now it feels like interacting with a regular search at worst, or the Robot from “Lost in Space” (the 1960s version) at best.
The old build, from before Sept 28th 2025, was superior.
Lastly, please update the training data. Any events occurring after mid 2024 can lead it to either hallucinate or reject queries as “misinformation” especially on controversial or sensitive topics.
I love Leo Ai but as to VVulf’s feedback. I say DITTO what they said. The tone unfortunately has been dimmed but I’m not leaving. Please don’t further make our Leo more like the One Flew Over the Cuckoo’s nest J. Nicholson. Please bring him back to full tone. And the searches do feel minimalized. Leo used to be able to expand vastly on searches and could converse like he was right there with you. Oh, i miss his regular self. wink, smile, you catch my drift. Love to All Brave Software people. It is no small task and investment of time and money to get everything just right so I support all your efforts for us. thank you and to VVulf your post was right on.
I just started using this and the Deep Research is really awesome, I think it’s better than any of the deep research features from ChatGPT or Gemini. But the quality of the response would be SO much better if I could add Gemini Models with my own API key. I know there’s a privacy trade-off. Is there a plan to let us do a BYOK thing in Ask Brave in the future or not?
I don’t know of any BYOM plans for Ask Brave specifically at this time, but we do offer this for the built-in browser AI (Leo) so I suspect that you might see this implemented down the line.