Qwen 3.8 27B: New FREE Local AI

Today we have the brand new release of probably the best free open source local model that you could actually use called Quen 3.8 27b. We'll talk about some of the best setups for this because again with local models we're always limited in terms of making sure you've got the right setup but basically this is an open weights version of Quen 3.8. So Quen 3.8 max is their frontier model. Quen 3.8 2.4 t a 95b has also been released that's a max version but the one that we're focusing on today is Quen 3.8 27.

Now if you're wondering how good is it basically if you look like we've got Claude Opus 4.6 that was around six months ago and that was considered probably the best model around and now you've got Quen 3.8 27b basically performing on the same level for free locally. What's your take Gazza? Yeah it's um I was looking at some of the stats before and you don't need a stupidly crazy graphics card to run it as well. Very interesting.

I almost feel like Quen has taken over the spotlight of being the go-to. Oh yeah I think you're right. I think you're right. I think they're leading in terms of that because like better than like you know you got glm 5.3 that's going to be open weights but you can't just run like you just need a ridiculous setup.

The same for example with uh Kimi K3 like it's open weights but realistically is anyone actually going to run that locally unless you're using something like Cerebri. Yeah yeah that's it and and also for the average person they wouldn't even know how to set it up like where to start or anything like that so I think with this setup if you had a dgx spark if you had a rtx something like that it's actually going to run pretty well. I think you're on a mac mini I'm on a mac studio I probably wouldn't run this I don't think it's going to run probably for me. Yeah so I've got a mac mini and I also have a pc with an rtx graph I think maybe my pc would be able to run it.

Mac mini with a macbook pro definitely. Yeah yeah that's it and usually when you try those models like Claude tells you you can run it and no problem mate and then you're in the middle of like uh you're in the middle of doing something around a meeting and your your whole setup just crashes uh completely. It reminds me of the days where like game developers would say oh you need eight gigabytes of ram 50 gigabyte storage space and then you would actually get the game and it's nah mate you need like 16 gig of 100 gig. Yeah yeah that's it isn't it it's like in reality you know it's it takes a lot more than you think but for example if you look at this as well I mean opus 4.6 max is not far away and this is what the top model probably six months ago if you went back.

Yeah it's unreal when you think about that. It's crazy just out of curiosity quen um is there like a is is that their best model 2.8 or is have they got like a better model? Yeah so quen 3.8 max is their best model they have uh released the open weights for that but it's like again you just wouldn't be able to to self-host a local model. Right okay yeah it's it's it is insane how much informants.

I think I said this on one of the other videos where the the guys that are currently at the top which is obviously Claude and Froppit could even say chat um it's very easy for the guys 5th 6th basically rip what the guys at the top are doing into that cap so it's it's really interesting to see competition point who's actually able what's that proving. Yeah 100% um also if you look at quen 3.8 so people will be wondering okay how do you set up so you could use like LM studio that's a free app and then you can get the model from LM studio as well and then also you've got olama so olama has quen 3.8 and it's just been updated 16 hours ago with the new update so you can actually one thing that's pretty good is you can get this on mox so if you're on a mac like mox basically means it will run smoother it's a better way to like run local models again 18 gigabyte model is that going to run smoothly on a mac studio probably not but you can set it up with claude code or you can use it as a brain for open code hermes asian or open core yeah out of curiosity which setup you reckon using the olama method or is there is that the best way to self-host i think so i think so like i think it's probably the easiest way because it's just like you click a button you paste that into your terminal and you get the model you can use LM studio i don't find LM studio it's nice to use there's something like yeah i've used olama for i think it might be the one thing i would recommend is these free self-hosted models sound great until you actually self-host and then realize how slow they are so yeah i've i've found an apple to get yeah that's it we actually we started testing out like local models on the leaderboards and we used to have it on on goldie bench and one thing we realized was like some of these models were just so bad running locally in terms of what they created and the quality of them i was just like mate we we just gotta stop like there's absolutely no point testing these models anymore because the locals end up it's so bad what was the worst model that you tested there was some that literally like they just wouldn't create anything actually useful and then you'd have to cycle through them again and again and again and iterate them i can't remember the worst one that i tried i mean we tried like ornith we tried stuff like quifos which was like supposedly like a mythos trained version of quen but honestly like i think you just need you know you just need a good setup expert you're good to go yeah so that's basically i mean some people say well why would you want to run it locally well basically it's free it's private you can run it offline if you have no wi-fi you could for example like use ai on a plane or whatever and also this is the frontier well not frontier but it's like it's getting towards frontier and also the gap it used to be like one or two years behind private and and locally hosted ai but it's like it's closing the gap faster and faster and faster and models frontier level are getting released at less frequent intervals i've seen that as well so it's kind of like i've i think you know give it six months give it a couple of years local models are going to be right up there it's going to be interesting more and more they actually end up optimizing it hopefully they end up optimizing it for low because the other the other side of the coin they might never actually end up optimizing use their server it will be interesting yeah 100 so thanks very much for watching if you haven't already check out the new search from casra dash this is his school community that basically teaches people seo how to rank with ai search engines he's built a awesome seo agentic operating system inside there as well and then if you haven't already feel free to check out the air profitable name as well these are both link in the comments description or go to the air profitable for this one and yeah you'll get our community and coaching and all our best trainings on this sort of stuff so cheers for watching see you on the next one