Leo AI Some Question Regarding Limit and Future of Leo

Browser Support Leo AI
1

Hello there, Brave Developer, i recently trialed the Leo AI, its genuinely pretty good as Browser Co-Helper althought some lack of feature kinda stop me from all in with my 20 bucks for starter

its a bummer persistent workspace is still in development or..its in backburner? i dont know but its apperently will come one day

Obviously, the brave developer cannot burn cash forever those API definitely are not exactly cheap nor is self hosting it, specially by offering Opus as part of the premium package, which raised some question

does Leo artificially limit stuff?

take for example Perplexity, they offered Opus, but the context size is abysmal at only 30K Token

Does Brave Leo have similiar limit or is the self hosting somehow fix all of it? I suspect the Open Weight AI like Llama wont have that problem but Sonnet and Opus im kinda skeptical? i heard in the past the massage limit is 2K Token But i havent tested that long massage yet

also regarding future

I know the Brave team probably got this question a lot, but well…what is the future regarding Leo? Workspace is certainly one, and its would be really good feature

For Context, im a writter from AO3, i have whole Md files about my story, Formatting Guide because those Html are pain if you format manually, and guide how the AI Behave, basically i have entire md library, for making a Co-Writter yes Co-Writter, i made the 95% of the Text, while AI hammered the details and checking up the grammar, those md files are just assurance that the AI wouldnt hallucinate

Leo AFAIK for now, didnt have this feature, nor did they have native md files as far as i know?

so there is that, Those Leo have artificial limit (be it upload limit,context limit and so on) and what does brave team have for future? (workspace seemingly being explored by the dev team)

For Now during this trial period, due to the lack of ability of using brave leo as co writter, for now im stuck with using Leo as helper to summarize Pharmacopeia and other Pharmacy Textbook directly, this is genuinely a godsend since i dont need to upload the pdf i can do it on the browser, copy relevant passage and boom answer, and its helped get around the many Pharmacy Jargon, its helped me as Pharmacy student a lot, althought i still hope for workspace to happened in this year so i can continue my hobby of writting

anyway cheers for the development team

2

Hey! Apologies for the delay in getting back to you - been a busy couple of days

its a bummer persistent workspace is still in development or..its in backburner? i dont know but its apperently will come one day

We kinda realised that adding Skills ended up covering a lot of the benefits of workspaces/projects, it’s still on our radar but the priority went down I’m afraid. We’re quite a small team compared to the big AI companies so (it sucks but) we do have to be very aggressive in prioritising particular things. Sorry about that :frowning:

Does Brave Leo have similiar limit or is the self hosting somehow fix all of it? I suspect the Open Weight AI like Llama wont have that problem but Sonnet and Opus im kinda skeptical? i heard in the past the massage limit is 2K Token But i havent tested that long massage yet

We do have limits on things like context length/rate of messages/length of conversations. This is where the upside of being quite small comes in though - we’re lucky in that we’re able to keep these quite high as overall volume to these models is fairly low. We take care of most of the LLM requests (free users/random stuff like searching or conversation titles/subagents etc) using self-hosted models to keep costs down further too. Naturally if costs/usage spikes up, we may have to lower the limits for certain models like sonnet/opus but for now we’re happy to keep them quite high to give a better experience

I know the Brave team probably got this question a lot, but well…what is the future regarding Leo? Workspace is certainly one, and its would be really good feature

Again this is something that changes quite aggressively, we’re just after wrapping up our big plan over the past ~3 months: ‘Leo v2’ which is in Nightly now - there were a few things which went into that:

  • Better self-hosted models
  • Subagents (improved searching, deepresearch which is imminent, etc)
  • Model-powered tool calling & server-side MCP client for adding new tools
  • Better alignment with BYOM API, so we can build things for both on the client side
  • Brave-trained LLMs (LoRAs) for specific tasks
  • Inline searches and richer outputs (graphs, code execution, image searching, etc)

And a bunch of other small stuff. Now that we’re putting the finishing touches on that, we’ve been putting together plans for the next big project. I shouldn’t speak for the other teams involved in Leo before they finalise things but I can give an idea on where we’re planning to go with the ML team - we’re going to lean more into subagents/orchestration and offer a more ‘dynamic’ Leo which uses smart triaging to use different tools/system prompts/subagents to let Leo. Essentially letting us add a TON of new tools and functionalities to Leo and making the experience on the whole ‘smarter’ without degrading overall quality by overloading the models

Leo AFAIK for now, didnt have this feature, nor did they have native md files as far as i know?

We do have Skills in the / menu where you can add custom instructions for specific tasks, markdown instructions should definitely work there. Not sure if I’m fully understanding the gap so it might not be useful, but does that help get you there?

1 Like