Which model to use for coding: qwen3.5 or qwen2.5-coder?

Posted by Mashic@reddit | LocalLLaMA | View on Reddit | 23 comments

I'm kind of new to local LLMs. I can see that Qwen offers dedicated models for coding qwen2.5-coder and they have the newer general models qwen3.5, should I use the old coding dedicated model or the new general one. I'm using them with VSCodium and ollama app.

[-]

LightCompetitive4266@reddit

People are literally dumb here. the question was qwen2.5-coder:14b vs qwen3.5:9b, and they are talking about 35B models. Like seriously.

[-]

zipeldiablo@reddit

There is still no answer :/

[-]

Pristine-Woodpecker@reddit

There will never be one. When OP asked his question, he didn't include this at all, so nobody answered it, of course.

[-]

Pristine-Woodpecker@reddit

He added that part in an edit after everyone had already responded.

[-]

Pristine-Woodpecker@reddit

Qwen2.5 is years old, it's multiple orders of magnitude worse than newer models. It should never ever even enter consideration.

Qwen3.5 (397B, 122B, 27B) is superior to Qwen3-Coder-Next as well. Only the 35B-A3B is slightly worse, but it's probably way easier to fit on your hardware.

[-]

Sharp_Classroom9686@reddit

No es lo mismo un modelo general que un coder especializado. Me imagino que en 9 meses las librerías y frameworks como React, NextJS, Django FastAPI ETC han cambiado tanto que el modelo ni las coreconoceria.

Utiliza una versión coder. Qwen3 o 2.4-14B. Utiliza Aider. si vas a utilizar OpenCode asegúrate de tener al menos 32K en contexto.

Y también utiliza el que más TK/s te esté dando a ver si terminas alguna parte del proyecto antes de que salgan los Qwen-6 🤣

[-]

Pristine-Woodpecker@reddit

This is outrageously bad advice and you should feel really bad for posting this.

[-]

Sharp_Classroom9686@reddit

This sounds less like engineering advice and more like shiny object syndrome mixed with benchmark addiction.

You talk like someone who hasn’t actually spent much time working with real code and local agents. Do your own testing. Run Gemma 4 26B A4B, Qwen3 Coder 30B, and even Gemma 31B if you want.

Yes, Gemma may give you more tokens per second, but it also tends to produce more stubs and half-finished code. Fast output is not the same thing as functional output.

[-]

Pristine-Woodpecker@reddit

Qwen3.5 is significantly better than Qwen3 Coder even for coding tasks, let alone 2.5 which is ancient and was never very good! It's not because it's not called Coder that it shouldn't be used for coding. The opposite in this case.

Gemma 26B was completely unusable in all my tests and I stopped trying to use it for production work.

This sounds less like engineering advice and more like shiny object syndrome mixed with benchmark addiction...You talk like someone who hasn’t actually spent much time working with real code and local agents.

🤣🤣🤣🤣

[-]

Old_Hospital_934@reddit

If you system allows it:
- Either Qwen3.5 122b-a10b
- or Qwen3 Coder 80b-a3b

If your machine can't handle that (given that you have a normal pc like the rest of us), qwen3.5 small models/medium models are just as fine. 35b-a3b is really good.

[-]

sanjxz54@reddit

How much shared is needed for a10b q4? I tried to run it on 12gb vram+56 ram and it didn't work . Should I upgrade to 80 or 96 ram and what can I expect on 5070 and ddr4 3600+5700x3d? I daily a3b right now and like it so far after tool calling fix, but really wanna try a10b. Or should I get coder next 80b?

[-]

Old_Hospital_934@reddit

For memory, I can say. the 4 bit ver. of 122b-a10b is \~65 GB, and the memory for \~256k tokens should be \~10GB or less (thanks to the new hybrid architecture)

So yea, around 80 Gb ram should do, get 96 for stability.

From what i've heard, coder next 80b is also a great general model, that should do it for you.
I'm not sure about the tokens/s because I don't have the hardware to run either of those, I cap out at 16 gigs😓

[-]

Pristine-Woodpecker@reddit

The 27B without thinking can also be interesting, it can beat even the 122B.

[-]

Old_Hospital_934@reddit

Yes, dense will almost always beat MoE, in my observations if the activated parameters is less than half of the dense model, it beats them. Note that this is only for Medium models. Ultra Sparse (500b and above) like Ling2.5, Ring2.5 and Kimi k2.5, GLM 5.0, etc. all beat. it.

That said, I absolutely love the 27b, I would def. upgrade my setup to run it as a daily driver (for heavy tasks)

[-]

EbbNorth7735@reddit

Technically both 122B and 80B are medium models. I feel like an enthusiast rig maxs out at around 48GB of VRAM. Two 3090's, and 64GB of system RAM sprt of level. If you're lucky enough to have more that's awesome but not required. I think both of those models would fit in 112GB of combined RAM.

[-]