Coding comparison between the Gemma 4 and Qwen 3.6

>On a my 16vram video card. Both models runs comparable speed. >On Windows LM Studio using recommended inference settings. >Model used:
>unsloth/gemma-4-26B-A4B-it-UD-Q4_K_S

>AesSedai/Qwen3.6-35B-A3B IQ4_XS

This guy says the speed is the same. What the fuck.
There are a bunch of interesting comments on how to improve Gemma output. Apparently it's not human-preference tuned out of the box, but you can put some hints in the system prompt and it get's cracking. That's the third pic. Yet others say Qwen also gets better if you do this.

At this point I need to find or build a benchmarking pipeline. I am not doing this manually every time a new model comes out!

via https://www.reddit.com/r/LocalLLaMA/comments/1sqxiz0

RT: https://poa.st/objects/662100a6-ccc3-4076-878a-c26ad20453f3

@WandererUber THANK YOU CHINA

@bronze @WandererUber i'm kinda worried because last i heard qwen is kind of on a rough spot right now. I root for them though, I use Qwen3-TTS for my AI dubbing workflow.

>my AI dubbing workflow
link me

@WandererUber @bronze It's tied to my IRL identity, but I can tell you the software I use:
- whisperx for subtitle generation
- gemma3n (when local) to translate the subtitles
- extra software to handle cutting up the mp3 file to have the voice segments

Then I pass the audio segments to koboldcpp running with qwen3-tts for it to clone. Gets the original audio, the text, spits out the result. It's not perfect, but it lets me watch a german TV show with mum in spanish.

The support for qwen3-tts isn't 100% wired though, I can't specify the language of the output audio, so sometimes I get french, german or english voices speaking spanish.

NGL being able to do this on a little iGPU makes me happy but also annoyed at how classmates at uni barely use their beefy gaming laptops for gayming and even then...

Really really cool. Thank you.
>iGPU
We've come such a long way, haven't we.

fwiw GLM 5.1 is free on modal for the rest of the month, so if you need a powerful model, try that. Lain got me onto this.

@WandererUber @bronze I know of it, but not too thrilled due to the sword of damocles hanging on that :/

I aim to someday have the money to get a gaming laptop or a macbook with enough ram for AI stuff + all the other crazy shit I do.

@WandererUber @bronze also idrk if you'll care but I also generate anime women.

@mischievoustomato @bronze @WandererUber
wood.jpg

@j @bronze @WandererUber what i'd do if she was real and mine, man