In local AI coding tasks, larger language models (35B parameters) typically generate code faster (146 tokens/second vs 43 tokens/second) but may produce less functional code, while smaller models (27B parameters) generate slower but more reliable and complete solutions, demonstrating a fundamental trade-off between generation speed and code quality in LLM-based development.
深掘り
前提条件
- データがありません。
次のステップ
- データがありません。
深掘り
Qwopus3.6 35B A3B MTP vs 27B MTP | Local AI Head-to-Head追加:
Smooth landing, I'd say. I'd say so.
My god, that is Oh, we are chasing some tokens today.
We've got Corpus 3.6 27B MTP and we've got Corpus 3.6 35B A3B MTP.
I want to see that speed. I want all of it.
We're using open code web UI side-by-side same time same prompts.
We've got them running on llama CPP and we're using llama swap to get us going here. The models are already loaded in.
You can see by my head we are filling up that VRAM to the max. Now, they're different quants, but this is what we could fit on this these 24 GB cards. So, if you're in this situation with a 24 GB card, this is what you're going to fit. Now, you might want to step it down a quant because I can only fit around 40 50k uh context.
Maybe we can push it a little higher, but I will I always like to leave some headroom. So, we're going to do three prompts. First one's going to be a AI server fleet control panel. I want to see the control panels. I always I want to see that dashboard, baby. Give me that sweet dashboard. Then, we're going to go with a galactic colony survival game and then we're going to do a monster truck mayhem game.
We'll see.
I don't Are they going to be too difficult for 35B?
That's what we're going to find out here. All right, so we've got the seeds locked for MTP. We've got the precise coding parameters in that config for the quant models just like they like it. And first prompt, we're going to do the AI fleet control panel.
Let's throw it in here telling it it's in a head-to-head coding competition with another model.
Check over your work plenty of instructions and directions here.
Don't give me any errors and output to a single HTML file. All right, after it's done, we will check the time, the tokens per second, and the code output. And if y'all want to follow along, tokenchaser.net, you can find all of the videos with all of the prompts and all of the code outputs. You can view them, check them out, grab the prompts down in your own models, see what happens. So, let's go.
All right, both are done here.
Looks like they both went back over, skimmed through their work.
27B made a correction or two.
So, 27B coming in at 9 minutes 9 seconds.
35B coming in at 2 minutes 47 seconds.
The file size is 27B came in at 73 kilobytes, and 35B came in at 65 kilobytes. All right.
And tokens per second.
27B came in at 43 tokens per second. Pretty good. MTP going to work there.
And 35B coming in 146 tokens per second. God, that's beautiful.
All right.
Let's check out 27B's first.
That looks good.
I like that.
And we're not going to spend too much time on these dashboards here. If you want to check them out, they'll be on tokenchaser.net.
Ooh.
I love when it puts in all of the pages.
Don't give me this nice sidebar menu with only one working page. It filled up everything.
Refresh works.
Search.
Search works. Very good. Very good. It's very detailed. Looks like it expands off of the page here.
That that looks good.
I don't see any errors. All right. Very good. Let's check out 35B.
All right. All right.
This is an interesting design here.
I think I'd prefer the left side menu.
But hey, it works. It it's good.
I don't see any errors.
These are refreshing, pause, resume.
I mean, that looks good.
Can you search?
You can. Search works.
On each refresh, they're refilling the whole bar. All right.
Get out of here.
Very good. Dismissed aren't they grayed out?
These graphs do look nice.
They're updating live activity log.
Very good. Hey, that's good job for a one shot on that 35B.
All right. I like it.
Next prompt. All right. Next prompt we're going to do here is a galactic colony survival game.
Lots of instructions.
Uh it's it's a strategy type game.
Um you're going to coding head-to-head.
Check over your work. Don't give me any errors. All right, let's see what happens.
All right, both are done here.
27 B coming in at 5 minutes 36 seconds.
Uh 35B actually hit the context window max. I will go find how long the first run took and put it on the screen there, but it automatically compacted and kept going.
So, that was great to see it do what it's supposed to do. I'll probably bump up the context a little bit for the next run here. Let's see the 27B model coming in at 38.12 tokens a second and 35B coming in at 169.86 tokens a sec. God, that is beautiful.
We can see down here. See, it hit that 40K window pretty quick, actually. I wasn't expecting that, but then auto compacted and kept pushing along. So, that's good stuff. All right, let's check out the 27B Galactic Survival first.
All right. To be honest, I don't know what I'm doing here, but I'm doing it.
All right, so this is quite a uh strategy type game. It's uh pretty cool. Uh it seems like all the logic is there. We've got the dust storms coming by. Um you can upgrade things.
There's um supply crates coming down.
Seems pretty in-depth and it functions.
I haven't hit any errors yet, but this seems like it could take a little bit. So, as far as the game functionality, seems like it is working.
Got more supply drops incoming.
We've got the log down there.
The days are continuing.
That's pretty cool.
If you want to play that one or check it out, make sure it works up to day 30.
tokenchaser.net, you can check it out on there. But, I want to see if 35B came through with anything like 27B was able to do here.
Let's open up the 35B.
All right.
Okay. Well, right off the bat, I mean, it has a different UI for sure.
The 35B seems to have more functionality.
Everything has to It is um very fast-paced and I can barely click before I lose.
I mean, it works.
If you want to spend the time to figure out exactly what the hell you're doing, we've got solar flares, alien raids, dust storms.
Everything has tooltips, which is pretty cool. I think it has more features.
Does it look as good?
Maybe.
I think that's personal preference.
Got a dust storm coming in. It is just wiping out all my folks in my colony here.
If it could slow down a little bit, give me some more time to make some upgrades, figure out how to survive.
Got a supply drop. There's definitely more going on in this one.
Which is cool to see 35B do that.
Uh let's switch back real quick.
27B.
It seems a little simpler, cleaner for sure.
I think cleaner.
A little slower-paced.
Everything doesn't have tooltips. 35B mineral vein. I mean, this has a lot more going on.
It could probably use a couple fixes.
But, that is cool to see.
A lot of pop-ups.
It's It's very fast-paced.
We need to slow it down, but all right.
That's good to see. They both completed it.
No errors first go.
They both function. Food surplus.
It has a lot of features.
That's cool. All right.
Let's go to that last prompt. All right, last prompt we're going to do here is the Monster Truck Mayhem.
I hope they go hard on this one.
Lots of instruction.
It's a side view type game. Go over your work. Don't give me any errors. All right, let's see what happens.
All right, both are done here. 27B ran into a little issue. I did push the context window to 50K. It seems like that may have been a little too high for the 27B model with this amount of VRAM because it actually crashed llama swap uh when it filled the VRAM. It I think it got to like 49K cash and then uh llama swap crashed, restarted, compacted, and it did finish.
So, everything is still good, but that's just something to think about.
Um I'm not sure how long it took. I'll put that on screen there and 35B coming in at 2 minutes and 28 seconds. The tokens per second this isn't really going to be uh that you know, 43.38. All right, we'll call it that for the 27B and 35B 145.08.
That is good stuff. Okay, 35B actually for some reason did not save the file.
It saved it in a temp directory. I had to tell it I don't see it and it output the file.
All right, so both of the file sizes are coming in at 49 kilobytes each. So, that's cool. All right, let's check out 27B first.
Okay, let's see. Oh.
Oh, we got an error.
All right, let's give it the opportunity to fix that.
And we will check out 35b.
Press enter. Press enter.
Wow.
Okay.
I got the flips going.
Not much else going on here. I'm out of fuel.
You can just rotate.
Okay.
Um let me give uh 35b an opportunity to make this a little better. We'll see what happens.
27b is ready.
This is This is something.
Can we Okay. I I Hey.
We're moving.
We're collecting coins. We've got other cars we are demolishing. We've got boost.
We can lean forward and backwards.
There's no way we'll be able to do a backflip with the sensitivity.
The wheels are not touching the truck.
Whatever.
We've got sound effects. We've got working fuel.
How do you boost?
The The wheels spin. We've got headlight glow there. We've got smoke coming out the back.
I mean, for a one-shot, okay, boost works.
Mother of Christ.
Holy All right, boost auto regens.
We've We've our HP up there.
All right, let's pick up some speed again.
My god, that is Okay, this is playable.
Oh.
Okay.
That's interesting. That I mean, if I could do some flips, that would be cool, but it functions and it works.
And that was while uh llama wob crashed because the contacts window, it rebooted itself, auto compacted, and finished.
That's cool. All right, let's see has 35B got a working product for us here.
What's going on?
Where we What are we doing?
Okay.
All right, 35B does not have a working product, but I've got an idea. I want to try something. I'm going to allow 27B to go look over this 35B uh monster truck file and make it a functioning game with all of those requirements. So, let's try that.
All right, let's see. Simple prompt.
All right, 27B is done attempting to fix the 35B monster truck mayhem.
Let's see what happens. It did have to go through and compact contacts once.
Uh it might be worth going down a quant size and stretching that contacts window much higher so we don't have to compact so often, but hey, if you want max quality, it's a it's the trade-off you got to figure out for yourself, but all right, let's see what uh the 35B monster looks like now.
Is it playable?
Okay.
All right, looks like we are all done with the jumps already.
We're just cruising. Let's restart that.
Hey, that's a flip, my friend. Couldn't flip in the other one.
Smooth landing, I'd say. I'd say so.
All right, so it's It's um Hold on.
Hold on.
It likes to wheelie by itself. It's putting everything right there at the start.
It's got some spikes.
It's got boost pads, ramps, coins, a car to destroy.
It has everything. It's all just compacted right there at the start. So, uh I'm sure a couple more prompts would would get us get it going just fine.
But, I I could see that 27B took much more time to do these fixes and more contacts than just creating its own new file, probably with additional features. So, that's something to think about. The We've got some either smoke or dust trail behind us.
We can jump.
We got flips.
Okay. I mean, okay.
So, it's getting there. Probably another go um and we tell it what's wrong, it >> [music] >> it probably could produce something decently playable. But, 27B's is going to take the win on that one. So, overall, 27B for coding is the go-to.
But, that 35B, A3B, MTP moves quick.
But, that's the trade-off.
All right. Well, you guys let me know what you think.
Who won? What should we do next?
Uh I do like this MTP. I think I'm going to start daily driving it. So, that is it for today. Thanks for watching, guys. We'll see you in the next one.
関連おすすめ
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











