Devstral Small 2 24B vs Qwen 3.6 27b or both? 1x 3090

Posted by szansky@reddit | LocalLLaMA | View on Reddit | 14 comments

Hi got 1x 3090 and I'm thinking about these both models. I'm using from Friday Qwen and this model is amazing! But.. what about Devstral Small 2 (24B)? Worth? Or not ?

Reply to Post

14 Comments

[-]

INT_21h@reddit

Qwen 27B is newer and, by the numbers, substantially stronger. I used Devstral 2 Small as my primary model for a long time, and what I'll say about it was it made a great *collaborator*. Wicked fast prompt processing, and no reasoning trace, meant never waiting long for a response. And where a confused Qwen devolves into overthinking and obsessively combing the whole codebase for clues, a confused Devstral devolves into "French nonchalance" returning a lackluster answer much more quickly. This meant it was often much faster to point Devstral in the right direction and get good answers. However, I don't run Devstral any longer. Gemma 26B-A4B has taken over its "fast idiot" niche for me. You might want to consider that one as a partner to Qwen.

[-]

Jorlen@reddit

I realize this post is a month old - but I'm still learning about LLMs. I just had a quick question for you. Since you like Q3.6 - 27B (and why wouldn't you) - why not go for 35B-A3B (also in q3.6 family) for the "fast idiot"? I am just curious as to what you prefer from Gemma's A4B vs the A3B. Do you find it does a better job overall?

[-]

Ok-Internal9317@reddit

A single 3090? I thought you would struggle in prompt processing what is your speed and quant?

[-]

PhantomGaming27249@reddit

Qwen throws hands with sonnet max, it outclasses basically anything less than Kimi 2.6/glm5.1 or proprietary models that need multiple extremely high end gpus.

[-]

curios-al@reddit

Both are solid.

[-]

HopePupal@reddit

every Devstral has been a disappointment. that one is no exception

[-]

StardockEngineer@reddit

Nah Devstral was the king for a bit. Punched far above its weight and did a lot of solid work for me. Especially as the subagent for a frontier model. Saved me a lot of money.

[-]

Nindaleth@reddit

In my very narrow evals Qwen 3.6 35B-A3B ended up stronger than Devstral Small 2 24B. Since the 27B Qwen should be even stronger than the 35B according to the popular opinion, Devstral Small would likely lose.

[-]

Due-Function-4877@reddit

Qwen3.6 27b a3 has been an incremental improvement at the cost of speed for me. I've had okay luck trying q4 kv with Qwen35-27-a3's as well, but it obviously degrades the outputs. I remember Devstral-2 was another one that tolerated q4 reasonably well compared to other models, but still lost quite a bit.

[-]