Mixtral - Best Opensource model broken down

Key Takeaways

how we doing today how we doing F so today we got a new open source model released um it's the mixture of experts a high quality sparse mixture of experts it's from mistol AI we know about mistol they dropped the 7 ballon parameter model that did exceptionally well using some lowkey methods that uh um are built I think on top of llama so um as you can see here it's gracefully handles a context of 32k tokens this new model um it handles English French Italian German and Spanish it shows strong performance in code generation more so than cold llama so that's a big development it can be fine- tuned into an instruction model that achieves the scoree of 8.3 and Mt bench and Mr already released her instruction model of fine tun in one and as you can see here it out competes GPT 3.5 on pretty much every Benchmark and of course llama 2 on pretty much every M Mark as well so this is like the best open source model today so far um they use some complicated methods it's called Mo um I don't want to look in too much into it because let's leave it to the exports I'm just going to use the model um that's what I do I use the tools that they make um I create tools but I use the open source models that other people create because I am not an expert in creating llms so let's start by trying out this new model so to do so I'm using together. a I'm not sponsored this is just the best platform I found um and you just make an account you log in and now I'm in and then they have all these other models that you can choose from but today we're just going to test out M mix 8 * 7B instruct so technically they have 32 billion parameters or 34 I think something like that but they have a method so it only uses 7 billion to generate which it which results in very quick generation like 100 tokens per second it's like cheap five level speed and yeah so let's try it out so we can have the output length at whatever tokens we want I'll put around 158 doesn't matter lower temperature a bit Let's test it out create a snake game in Python boom so using the curses Library let's try this out and then we'll compare it to 3.5 so let me open up chat gbt just for this comparison make a new chat I'm got I'm using 3.5 I never used 3.5 in a long time um create a snake game in Python and then we're just going to compare the results that's what we're doing this video so bum let's see if it fails let's see how good it is we going to copy code I'm going to go to python so let's copy this we'll name this mix and then we'll just play this but first we're going to install the dependencies that we need to install oh it looks like we don't even need to install anything so we'll just run this now mixp uh what happened o what happened here looks like I failed already what a failure let's try chat gbt looks like they did it properly I'm going to try again I'm going to give it the benefit of the doubt and try one more time because I that's not anywhere close to what we wanted so we'll just name this 53 paste this what do we have to install P game I think I'm going to prompt it to use P game for mix trol it might be cheating but gb5 is also much more fine-tuned than um yeah so we got us like a decent snake game here so let's tell it to make a so we're going to brace length before trying again make a snake game in P game and outp put and give it in one full one full code five okay looks like a p game now it's a little bit of a cheating but you have to understand gbt 3.5 is obviously more fine-tuned so as long as we prompt this correctly it should hopefully make a better game don't what 3.5 did let's run it we got the code let's copy paste run okay oh interesting so okay they have a restart screen this is better than 3.5 did right to make it fair I'm going to reprompt 3.5 and see it okay it worked okay it actually works quite well let me reprompt 3.5 turbo to give it a fair shot so going to restart let's see what happens awesome let's try again now let's try the open AI version ooh now he put a scoreboard okay to be honest this is much better than the mixture one but it crashes after running so I guess mixl did a restart screen that's I think all I'm going to test it for now um this model is good if you want to use something open source even though it's pretty much on bar with gbt 3.5 it's not gbt 4 level or anywhere there yet um but you can see this is an open source model and things are just getting better and better and better so stay tuned for the next video maybe I can show you guys how to implement it in a in an app because I'm currently using 2.5 for my tools but I'm going to start probably integrating it with uh open source models like Mixel less restricted and more freedom so even together AI they let you import and use um Mixr through their API so you can use it that way you don't have to host it locally peace