DeepSeek-V4.1-Flash-1017 on a single AMD MI325X
4 pi agents at 100 tok/s
Native MTP + mixed quantization at near-lossless accuracy
github.com/lxe/dsx/blob/m…
This is the latest iteration of my setup for the Muse Spark Open Kernel Challenge.
The earlier versions ran a small team of long-lived Claudes that decohered after 12 or so hours without hand holding.
I replaced this smorgasbord team with a thin harness. A swarm of 6 (for now)
I've learned so much since this tweet. In general, this has been validated.
If you are doing a hybrid CPU and GPU approach to running local models, you really, really need to pay attention to what workloads and hardware fit what kind of quantization techniques.
