Qwen 3.8 Flash-Next has been such an incredible release that I unfortunately could not realistically run on my setup (16GB VRAM + 32 GB RAM) so I wrote a custom inference engine and I can now get 10+ tok/s on UD-Q4-K-XL from @UnslothAI and 240+ tok/s prefill!!
@QwenDevs
@bnjmn_marie If your setup cannot accommodate and hold the full UD-Q4-K-XL weights in VRAM + RAM would love if you gave QwFNfer a try to see if it would speed up your workflow !
@Oluwaphilemon1 Could I tempt you in trying out QwFNfer? I made it specifically for when you have less overall ressources than the model size allowing people to run the UD-Q4-K-XL quant on 32 GB of RAM and 12 to 16GB of VRAM at 20+ tok/s of decode
https://t.co/V6At4futYo
@dahoodie@coldniko Would you be willing to try QwFNfer?
I don't have access to that much hardware and would love to see how it'll compare!
(no multi GPU support yet)
https://t.co/V6At4futYo
@DeepBlue9696@tirzahitonan@coldniko Would love for you take a look at QwFNfer!
Different architecture than Strata and targets the higher quants of Qwen3.8 Flash Next more specifically UD-Q4-K-XL!
@0xSpuddy@joshuajohnsonAI Don't wanna shamelessly self promote but if the model gets stuck in a loop, maybe try a higher quant with QwFNfer? I made that engine specifically for higher quants of Qwen3.8 Flash Next (UD-Q4-K-XL) with less RAM and VRAM usage
Quiet two weeks but QwFNfer is back in force! Windows version is now officially out!!!
Thanks to everybody that is contributing to the project and using it!
As a reminder, QwFNfer let's you run big quants of Qwen3.8 Flash Next on small hardware!
https://t.co/xoaEDkuXak
@Vault_Dweller31@ItsmeAjayKV If you would like to try, I'm developing an inference engine specifically for the qwen 4 architecture that has allowed me to run Qwen 3.8 Flash Next faster than llama.cpp on Linux.
If you want to take a look, here's the link to it
https://t.co/V6At4futYo
@net_termina Qwen3. 8 Flash-Next UD-Q4-K-XL with 16GB VRAM and about 20 GB of RAM.
I'm using a custom inference engine to get 15+tok/s
(if you ever want to try it: https://t.co/wZrHOQJH38)
@0xSero Don't wanna self promote but I made this and it allows people to run Qwen3. 8-Flash-Next at UD-Q4-K-XL at 10 tok/s sustained on long context and I'm still working on improving it!
Would love if you took a look at it!
https://t.co/wZrHOQJ9dA
@UnslothAI@QwenDevs Here is the link to it : https://t.co/wZrHOQJ9dA
It's still fresh out of the oven but I'm planning on optimizing and supporting.
It's made around the Qwen4 architecture so should work with the next generation of Qwen models too 😉
Qwen 3.8 Flash-Next has been such an incredible release that I unfortunately could not realistically run on my setup (16GB VRAM + 32 GB RAM) so I wrote a custom inference engine and I can now get 10+ tok/s on UD-Q4-K-XL from @UnslothAI and 240+ tok/s prefill!!
@QwenDevs
@Kyori_G@HelldiverGodGuy So you saw someone do an epic fail not researching anything about the topic and then decided not to look up what a designer does in a game studio... Good job but a designer is probably the reason why you like a game or not
@zemotion What about forensic watermarking on top of Glaze? Atleast you could trace back the scraping to accounts. Could help ban bot accounts and maybe also legally when pictures end up in datasets since you would have proof of the provenance of said picture
The kickstarter campaign is over and we reached €16,210 , which unlocks the ZIZAR MODE!! (coming a bit later after the release pls understand)
Thank u to everyone who pledged, this is crazy!!
Let's make a cool ass game now