I hate to be this guy but: Any good, recent CODING models in the 70-80B range?

Posted by ParaboloidalCrest@reddit | LocalLLaMA | View on Reddit | 22 comments

3x 24GB vram.
Qwen-coder-next is not bad. I'll continue to use it if you yell enough at me.
Larger than 80B and I'll have to sacrifice the decentish Q6 quant, or the minimum (for coding) 256k context.
I do NOT believe the latest 27-31B can realistically beat an 80B model, but change my mind.

[-]

wllmsaccnt@reddit

Qwen-coder-next is an 80B A3 (MoE) model, not an '80B model'. Quite different.

Yes, Qwen3.6 27B beats most other models up to 100B at coding benchmarks, and no it shouldn't be believable.

[-]

ParaboloidalCrest@reddit (OP)

You make it sound like QCN is 3B, which is not.

[-]

Thunderstarer@reddit

It is, though. In a very material sense, it is.

Total parameters are not equivalent to active parameters. Lacking in the former comes at the cost of knowledge. Lacking in the latter comes at the cost of reasoning capability. Qwen 3.6 27b is massively more capable of problem-solving than is QCN, but in exchange, it has less knowledge to draw from.

[-]

wllmsaccnt@reddit

I agree! It's not any more 3B than it is 80B. The MoE models are somewhere in between.

I say that as someone who loves the Qwen 3.6 - 35B A3 model (more because its fast than because of its capability, but its still my favorite).

Why not give them all a try, it doesn't cost anything.

[-]

Skelshy@reddit

Not that I know of. I tried Quen 3.5 122b and Quen3-coder-next as well as gpt-oss-120b. Always go back to 3.6.

[-]

grabber4321@reddit

GLM-4.5-air?

[-]

ttkciar@reddit

Still my favorite model, but a bit long in the tooth now, with no tool-calling skills to speak of.

[-]

grabber4321@reddit

I tried Qwen-coder-next, but for some reason it sucked - like the output was bad. Maybe its just my application.

27B still is very solid.

[-]

mukz_mckz@reddit

It's good for leet code like tasks, but sucks at agentic workflows

[-]

Juulk9087@reddit

3.5 122b sucks at nvfp4 as well.

[-]

Ok_Technology_5962@reddit

i think the issue here is q6. Dense models lose a lot more than MOE with quantization. try 27b bf16. Try that first at temp 0.1, top k 20 top p 0.95. Try without MTP as MTP also removes a bit of quality depending on what you are doing.

[-]

TimLikesAI@reddit

I cancelled my Claude Max subscription in favor of Qwen 3.6 27B on my GPU server w/ dual 5070 Ti 16GB cards. I have just enough room to squeeze a NVFP4 quant, q4_1 kv cache, and MTP.

[-]

huzbum@reddit

How well does the MTP work for you? Happy with it? I'm running 35b because I'm not patient enough for 27b, but maybe MTP could fix that?

[-]

TimLikesAI@reddit

Some raw logs from my current in-flight session: https://pastebin.com/GdhaFxgh

[-]

EvilGuy@reddit

Do you really find these models handle 256k context well? Anytime I let Qwen 27b go past around 100k context or so it starts to get loopy and the prompt processing really starts to be slow.

[-]

huzbum@reddit

I'm not sure they're wrong. Mine seems fine right up till it hits compaction at 120k-ish tokens and compacts. Then it's dumb as shit. Even after I get it back on the same page it's just confused.

[-]

Automatic-Arm8153@reddit

Your not wrong. I think the people who claim there are no problems past 100k context aren’t doing much meaningful work.

Even with cloud models they are noticeably degraded at 200K plus.

[-]

huzbum@reddit

Qwen3.6 27b. Fight me!

Seriously though, Qwen3 Next Coder 80b is only 3b active params. 27b has literally 9x more params per token. It may be a tossup knowledge wise, but when it comes to intelligence... it's got 9x as many neurons firing for each token generated. According to the Qwen team, it beats the 122b param model, it's definitely better than the 80b model.

I'm personally using Qwen3.6 35b (same 3b active params as Next, but less experts) because I prefer the speed of the MoE, even if it's not as good. Good enough for me.

[-]

Dany0@reddit

Q3.6 27B at fp16 with 1M ctx

finetune it on your own dataset and to your own benchmarks

[-]