This comparison proves that parameter count is often a vanity metric, as smaller, well-optimized models can deliver superior logic and fewer errors. It’s a sharp reminder that in local AI, architectural efficiency consistently beats raw scale.
深度探索
先修知识
- 暂无数据。
后续步骤
- 暂无数据。
深度探索
Qwen3.6 27B vs 35B A3B on RTX 3090s | Local AI Head-to-Head本站添加:
Damn, that looks nice. Go. Okay, start menu is I've seen better.
Ladies and gentlemen, Quinn 3.6 is out here making absolute waves in the LLM community. And today we're going to test head-to-head, side by side at the same time, uh, the 27B model and the 35B A3B model. Um, Quinn 3.6. Here we're running the thinking mode for precise coding tasks.
Highly recommend it. If you are coding with these models, definitely slap that in the config. We're running it on Llama CPP with Llama Swap to get us going here.
Uh, we'll be um firing off these prompts in Open Web UI. Just makes it easier to look at and do both at the same time.
here. Right above my head, you'll be able to see we're running it on a 3090 Ti and a 3090 locally. Uh, one model on each.
Uh, let's go ahead and get these models spun up here.
All right, we've got them running. You can see um 21.8 GB in the TI and 21.2 GB in the 3090. Um, we're using Quint 27 Quant 5 for the 27B model and quant 4 for the 35B model. It was really just whatever we could fit on that specific card with some decent context.
Uh, so let's go ahead and throw this first prompt in here. Create a single file HTML CSS JavaScript Windows style desktop with six apps, two of which are a task manager and a game. Nice and simple.
We'll see how they do head-to-head, how long they take and the outcome.
35B model is done.
27B model is just trucking along.
Finished at 1,300 lines.
All right, finally taking four or five times as long and less lines. We got 1181 coming out of that one. So tokens per second we're running around 33 for 27B model and around 137 on the 35B model. That is quite a difference. All right, let's get the 27B code opened up here.
See what we got first.
All right.
No rightclick menu. Interesting.
Didn't ask for it, but All right. Task manager looking pretty good. Let's see if Okay, applications do load in there.
And all closes everything. Very good.
No right clicking anywhere. Okay.
Quite a settings menu we got going on here.
All right.
Okay. Okay.
Lot going on.
Brightness works.
Okay, foul explorer.
Very good.
No right clicks anywhere.
Calculator works.
Notepad saved in there. All right. New Minimize.
Pull it back up. Good.
Cool. Works.
Snake.
Good looking snake.
Game over. It works.
Very good.
I like this task manager.
All right. Very good.
Let's check out the 35B model. Here we go. Okay. Start menu is I've seen better for sure.
Interesting.
Search doesn't work. There is a right click menu in this one though.
Task manager. Oh, interesting looking.
Cannot resize the windows.
Wow.
Okay, let's see.
I don't see the app actually popping up.
All right.
Check the calculator. I'm sure it works.
Cool.
Snake.
Not as good of a snake for sure, but it works.
Interesting.
There's no game over overlay.
It just stops. Waits for you to reset.
Okay.
File Explorer. Wow. Okay. This is actually the closest file explorer I've seen to a Windows one as far as the looks go.
But nothing works.
Can't click anything.
All right. Well, huh.
That works. Very good. Nice little terminal.
They went off with these the settings windows. Ah, well, of course, none of this stuff works.
That's fine. No brightness.
Nothing actually useful in here. No.
Change wallpaper doesn't work.
Refresh. Okay. Let me see. Minimize.
All right.
Okay, none of the buttons work.
The window itself looks very similar to Windows. That's pretty good. All right, cool.
Well, you can see the difference in the two. Coin 27 definitely has a better one, but took four times as long for a very two sentence simple prompt. Not much instruction. Just leaving it to figure it out.
All right, next prompt we're going to go with here.
Create a single file HTML CSS JavaScript cyber punk dashboard with neon panels, animated system stats, a live log terminal, network graph, yada yada. You can read it. Very straightforward, simple. We'll see the comparison. Let's go.
All right, 27B is finally done.
How many tokens a second did we come in at?
So, 140 for the 35B, 33ish for the 27B. All right, let's take a look at the 27V model first.
Damn.
That looks nice.
That's impressive.
That's got a lot of cyber punk going on.
All right, what do we got here?
CPU usage, memory, network, storage, GPU.
Interesting.
We've got a active stream of logs.
You can type. Does nothing, but all right. We've not really got any anything clickable, any popups, any active notifications.
Okay. Other than the fact it looks sick, this is something I'd love to build out into a useful dashboard of some kind. As far as the aesthetics go, that's that looks nice.
Nexus. Can't really read the Nexus up there, but All right. Very good. Very good.
Let's check out the 35B models code here.
Okay. 35B, what are you doing? What's going on with your banner situation there? Is that on purpose? Can't be.
Can't read. What?
I doubt it is on purpose. Is it?
Anyways, yeah, we've got some text behind it as well.
Behind. What the hell's going on here?
All right. I mean, it looks good besides that one little error.
Looks pretty nice. No worktopology.
Lots going on there.
We've got no popups, no nothing actually clickable.
We didn't tell it to do any of that. Um, gave it a lot of room to do whatever it wanted for the most part.
Looks very good though.
Very clean besides the banner.
Got the hover effects on everything.
Got the live system logs.
Can't type anything in this one. Looks good though. Looks very well. Very well done. Clean.
Fix the freaking banner. Come on. All right. Well, I'd say 27B one again. Now, the 27B model is running the Quant 5, which is going to be better than the Quant 4 a little. Uh, we were just fitting the biggest model we could on each GPU per model.
You know, if you're going quality, definitely 27B. If you don't mind this the taking four times as long, 27B is definitely the way to go. 35B still gets the job done. If you can run Q5, we get a little better quality, but the speed is unbelievable for what it's actually able to produce.
So, head-to-head coin 27B and 35B.
Both fantastic models. Both very well capable for coding tasks. But head-to-head, there we go. We'll see you in the next one.
相关推荐
OpenHuman VS Hermes AI: Who Wins?
JulianGoldieSEO
285 views•2026-05-29
Long-Running Agents — Build an Agent That Never Forgets with Google ADK
suryakunju
142 views•2026-05-30
This computer is made from real human brain cells. And you can buy it.
Talktmsmedia
3K views•2026-05-28
BREAKING: Microsoft’s New Image Generating Model Beat Out GPT 1.5 and Nano Banana 2
aimmediahouse
122 views•2026-06-03
I Made the Same Anime Fight Scene in Every AI Video Generator
NobleGooseAnime
295 views•2026-05-30
Nvidia Bets Big On AI PCs | New Chip To Power Windows Laptops | Technology | AI Updates | N18S
cnnnews18
3K views•2026-06-01
I Tested NEW Opus 4.8 on Four Projects (Updated LLM Leaderboard)
AICodingDaily
298 views•2026-05-29
3D Platformer Update - NO CAPES
SolarLune
294 views•2026-05-30











