alt Hacker News small native inference engine optimized first for DeepSeek V4 Flash (including the experimental vision model), DeepSeek V4.1 Flash (Metal, and text inference on CUDA), and additionally GLM 5.2 and 5.3, GLM 5.3 Flash and DeepSeek V4 PRO, and Qwen3.8 Flash Next (Metal and CUDA)
For our mere mortals that were kids not long ago and can't really believe we've got our hands on a x090 series targeting Qwen3.8 27b, https://github.com/noonghunna/club-3090 is the way to go.
I'm maintaining a web frontend for this, trying to at least. You can follow it here: https://github.com/gchamon/club-3090-server
First of all, a x090 series card is "high end hardware" in its own right these days. Secondly, I'd think you'd probably get more interesting results running a MoE model in CPU-MoE mode, i.e. with the shared parameters residing on GPU and sparse experts on CPU plus SSD offload. Yes it will be slower, but small dense models are just a dead end and not that interesting. (Note that prefill would still be sped up in this setting; the CPU/GPU layer split in llama.cpp and the like applies to decode, but even a "0 graphic layers" setup does accelerate prefill.)