Brave Leo defaults to Mixtral 8x7B

If you want an in-browser AI that doesn't hoard your chat history for model training, Brave just pushed a major upgrade to its native assistant.

Brave announced in its desktop v1.62 update that it has made Mistral AI's open-source Mixtral 8x7B the default engine behind Brave Leo, its privacy-focused browser assistant. The upgrade applies to both free users and $15-per-month Premium subscribers.

Mixtral replaces Meta's Llama 2 13B as the default choice. According to LMSYS Chatbot Arena data cited by Brave, Mixtral 8x7B outperforms ChatGPT 3.5, Claude Instant, and Llama 2, while also showing reduced hallucinations and biases on the BBQ benchmark. It brings multilingual support across English, French, German, Italian, and Spanish, along with beefed-up code generation and larger context windows.

Rate limits and model switching

While Mixtral 8x7B is now the default, Leo isn't locking users into a single brain. Free and paying users can still manually switch between Mixtral, Anthropic's Claude Instant, and Llama 2 13B.

There is a catch for non-paying users: free access to Mixtral and Claude Instant comes with strict rate limits. Once free users hit that ceiling, they can fall back to Llama 2, which has more generous limits, or pony up $15 a month for Premium to get higher quotas across the board.

Brave also polished Leo's UI in this release, tweaking onboarding flows, context controls, and response formatting. Desktop users can summon Leo directly from the address bar or via the sidebar, while mobile support for iOS and Android remains on the roadmap.

Decoupling payments from prompts

Privacy remains Leo's main selling point. Chats require no account or login, and all queries pass through an anonymizing reverse proxy so user IP addresses can never be tied to specific prompts. Conversations are wiped immediately after generation and never used for training.

Even more notable is how Brave handles paying users. Premium subscribers receive unlinkable cryptographic tokens to validate their subscription. This means Brave's servers cannot correlate a user's billing email with the actual browsing sessions or prompts they run through Leo.

Why it matters

Most browser-based AI tools treat user data as an acceptable tradeoff for convenience, forcing logins and funneling inputs straight into proprietary training loops. Brave is proving that you can plug a genuinely competitive model into a browser sidebar without strip-mining user telemetry in the process.

We think backing Mixtral 8x7B is a shrewd tactical move. Rather than being hostage to closed-source API pricing, leveraging an open-weights MoE model gives Brave the performance of proprietary chatbots with much higher architectural flexibility. For developers and privacy-conscious users who want fast webpage summaries and inline coding help without leaving a digital footprint, Leo is worth taking for a spin right now.

The main friction point looks like the free-tier experience. Bumping into strict rate limits and getting kicked down to an older Llama 2 model makes for a jarring drop in response quality. Until mobile apps arrive to complete the ecosystem, Brave Leo remains an impressive desktop tool that still has to convince mainstream users that browser privacy is worth paying $15 a month for.