FRAMEWIREIndonesiaUpdated Aug 21Live wire
0:00 / 0:00

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM)

I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B

AlokAug 2121
0:00 / 0:00

Deployed qwen 3.8 27b with dflash2 on 2 rtx 5090s

Steady generation: ~220 tok/s Normal range: ~190–257 tok/s Best observed: ~343 tok/s sgl_project with dflash2 is blazing fast

RajuGangitlaAug 211
0:00 / 0:00

Qwen 3.8 27B can now run locally on RTX 4060 8GB?!

This is what made me stop scrolling. The 27B model which usually sounds "too big for a small GPU" can actually be forced to run on an RTX 4060 with only 8GB VRAM using a new quantization from Unsloth: IQ4_XS.

stellar protocolAug 211
0:00 / 0:00

Gemini 3.7 Flash vs Qwen 3.8 vs DeepSeek V4 Pro vs GLM 5.3

Same prompt: busy train interior to leaving the station Qwen 3.8 & GLM 5.3 look the best visually

Md Ismail Šojal 🕷️Aug 212