Hey guys, I tested both of them and feel Qwen3codernext is more useful, whats your thoughts? thanks for your sharing.
I started running some benchmarks, and Qwen 3.6 seems to come out on top on those:
I havenāt done much real-world testing though, because itās really hard to compare that way and I wanted to answer the same question you had.
Iām currently running the 27B dense model with dflash to see what different it makes to the speed (in theory the quality should be the same). I was surprised that in the benchmarks Iāve run, the dense model hasnāt actually been as much slower as Iād expected (presumably itās more efficient and getting the answer and doesnāt go in loops or fail tool calls as often).
I find myself using one, then finding it is getting stuck in a rut, and then swapping out for the other to fix a problem, then swapping back again once that particular problem is fixed. It is increasingly looking like I need a cluster to run more than one at a time, and then use a smaller model to act as intermediary to the other coding modelsā¦
Problem is - things are moving so fast, new models seem to arrive every few days, and runtime recipes with different options seem to float around every few hours - it is difficult to keep track, especially as Iām pretty new to this.
I asked this question to myself multiple times. For my coding workflows, both work well and perform almost identically. I found myself going back and forth between these two and 3.5-122B-A10B (with all the optimizations from @Albond and @whpthomasā recipes it performs very well.
Sometimes ācoder-nextā get stuck in a few things and 122B fixes it at once.
Sometimes 122B loops itself and 35B-A3B-FP8 solves that at once.
I guess the real answer is āits dependsā :) Iām still defaulting back to 35B mosts of the times. Peforms very well and doesnāt make ugly mistakes too often.
We had good test bed with 21 KB, 30K documents, and really complicated requests like āFind one sided contracts focusing on termination termsā and so on.
We tested Qwen3.5 and Qwen3.6, Gemma4 and few more.
Qwen3.5-27b is 100% winner, with some tasks running better on Qwen3.6-27b. All other models we tested (Gemma4, NVIDIA models and so on) was very far behind. Qwen3.x need better tuning to work with complex file system tools and to manage context (looks as they were not trained to conserve context) and of course it all depends on a chat system a lot (we tested on Open-WebUI + RAG + Native function call so model calls RAG not chat) but results surprised us a lot . I run most models on fp8 quant on H100 or 2 x A6000/48.
And no, I do not see new models of the same quality since this time. Not yet. Iāll be glad to see similar models (or slightly bigger) trained better to deal with files and skills, but in general these are best open weight models for today.
I still have not found a model that fits on 2x sparks that beats qwen 3.6 27b 8 bit. Rock solid model.
Qwen3.6-27b performs very well, particularly in image recognition. however the token generation speed for the 27b model is quite slow, and has very little performance difference between single and dual dgx. and i also compared this to the Qwen3.6-35b model, I found that the 35b version often misreads numbers within images.
Both models were tested using MTP (native MTP for the 27b and DFlash for the 35b) within the Hermes agent. I am currently testing the 35b model without using MTP, as I heard that there has quality issues associated with mtp. If disabling MTP resolves the image accuracy issues, I will recommend the 35b model since 35b offers a superior balance of performance and quality, and very recommend with a dual node setup, as its performance scales linearly on dual dgx
hm for me 122b-A10B rocks
I tried the 35b on image captioning with no mtp and it did significantly worse than qwen3.6 27b. Both Q8 with f16 kv cache.
I ran a variety of benchmark tools and real-world, Hermes Agent-driven tasks against 27B and 35B. What I found was the benchmarks were close. 35B scored slightly better than 27B, but my opinion is benchmarks donāt reflect real-world use cases accurately, and, the difference could be noise.
Working on a complex coding project, both models generated code that was decent enough for first pass, in roughly comparable wall-clock time elapsed. But, 27B would frequently get in to loops when running tests. 35B would generally identify the cause of the failure, patch it, and move on without looping.
Both were better at real agentic tasks (including coding) than other models I tried, including 122B, Gemma and a couple of others. This was all on a single-node GX10, btw. Anecdotes arenāt data, YMMV, etc. etc.
Having used the Nvidia/Qwen3.6-35B-A3B-NVFP4 model without MTP for a few days, Iāve been sending it my Apple Health screenshots and smart scale readings via the Hermes agent. It has been tracking my records successfully, whereas it previously struggled to read the figures correctly when MTP was enabled. But lets test it for a longer period and see.