BasedBase

/

Qwen3-30B-A3B-Thinking-2507-Deepseek-v3.1-Distill

like 25

or even better q4 quants

I'd like to give this a try with other backends, would appreciate you uploading the unquantized files. Thank you so much!

Sure I will upload the unquantized model

or even better q4 quants

q4 quants and others will be posted soon

Thanks! I’ll quantize it to exl3 once they’re uploaded!

@BasedBase
Excellent will quant also ; and Brainstorm it.

Thanks! I’ll quantize it to exl3 once they’re uploaded!

@BasedBase
Excellent will quant also ; and Brainstorm it.

The unquantized model was uploaded

Thanks. Btw, have you seen kimi k2? Maybe it's worth doing distills based of that. This model is insane

Thanks. Btw, have you seen kimi k2? Maybe it's worth doing distills based of that. This model is insane

Potentially, I have 8TB of storage currently thats almost fully used so I need more storage lol. Also if I may ask, is my model the one thats "insane" or is kimi k2 the "insane" one. I need feedback on the models performance so I can accurately gauge which technique works the best for distillation and keep providing high performing distills.

Thank you!

I’d be interested in seeing how GLM 4.5 distills onto it

This looks awesome. Thanks a lot for helping out and doing this! I love researching this stuff. But quick question, do you did you use an imatrix for the quants? I'm excited, but I'll wait if you didn't.

This looks awesome. Thanks a lot for helping out and doing this! I love researching this stuff. But quick question, do you did you use an imatrix for the quants? I'm excited, but I'll wait if you didn't.

No I just used llama.cpp to quantize

The performance is exemplary! It reasons like a much bigger model. It is quite authoritative and follows instructions really well.

I gave it 13,000 lines of code in one prompt with max context of 262144. It took it all in, no problem. Addressed the issues I wanted without errors. If you want to see how analytical this model is, do the following: download a white paper off the internet and give it to the model. Ask it to improve the whitepaper. Sit back and behold the reasoning and critical thinking power of the model. I can hardly wait for future versions!