Betelgeuse running DeepSeek V4.1 Flash Q8_XL (1017) at 19 tok/s decode & 62 tok/s prefill on an RTX 3060 12GB + 16GB RAM: How we finally fixed the prefill bottleneck on out-of-core MoEs
After my last post on Qwen 3.8 Flash Next, a lot of people were asking if out-of-core MoE streaming could ever work on an actual frontier monster like DeepSeek V4.1 Flash (1017) , or if the prompt processing (PP) would make it completely unusable.