Qwen3.8-Flash-Next sounds like Claude and I hate it
Qwen3.6-27B and Qwen3.8-27B are great. They are fast and quite accurate for not having to spend too much time thinking or searching, as other models with comparable scores have to.
But Qwen3.8-Flash-Next, which tops a lot of benchmarks in its category, is the most obvious Claude distill ever, and although I don't particularly care about how models get trained, I hate the way it sounds and interacts with me.
- Three warts I'd still fix
- Statements are down to the two things code can't say
- The bug is fixed at the substrate that caused it
- Genuinely good now
- Also worth a conscious nod
- Say the word on 1 and/or 3 and I'll do them
Not to mention Claude's over-the-top comments and git commit messages that make you want to scream.
I wish Alibaba would go back to adopting its own style. Qwen3.6-27B and Qwen3.8-27B sound much better.
9 replies
I hate Claude's voice so much. So much ceremony spent on a few simple questions and caveats. I feel like AI defaults to assuming you have no idea what you're doing. ChatGPT is the same but with different verbal tics.
Let's cut down to the heart of the matter as you hit the nail squarely on the head
Never used Claude, but GPT-5.6 Sol annoyed me so much I actively avoid it and just use Terra on max. Sol is trained to favor convoluted San Francisco jargon that projects any trivial idea into some sort of an ingenious business maneuver.
I did notice the same. Both Sol and Claude prefer pushing that typical corpo business/product talk that makes me go crazy.
More specifically it keeps talking like that one colleague who speaks fluent corpo BS and manages to amaze non-tech people by verbose statements that ultimately say nothing or are super vague on details.
I only use Qwen for Zoo Code, when chatting I just use Gemma
That seems to be the consensus on Reddit, Qwen for code, Gemma for everything else. At least it was so for 3.6, not sure if 3.8 changes the situation.
I have to use Claude at work and have found myself wishing I could use Qwen 3.8 27B at work so I don’t have to waste time deciphering a machine’s fart sniffing. Guess it’s a good thing I don’t have enough VRAM to run a 125B model.
Yes, Opus/Fable 5. If you have the opportunity, play around with it. It writes in a format that tries so hard to be concise (while blurting out a monumental essay) that you can hardly understand it, and repeats the same phrases all the time.
I usually use skill like openspec when requesting LLM to do something, combining them with matt's skills like grill-me. I assume both of this skill will make LLM output in certain way as I jump between model, I don't really notice them. Usually I also ask to 'talk in basic english language' or ' talk in simplified technical english language' when just asking random things in the codebase. Surely not perfect and better to have normal sounding LLM directly.