DeepSeek-V4-Flash 284B on 5.3GB of memory

Following up on my Qwen 3.6 port, I wanted to keep adding models and ended up fixing a bunch of things along the way, so it's its own engine now: Mference. Same core idea from TurboFieldfare, MoE models activate a few B params per token, so keep the shared core and KV cache resident and stream

中文全文翻译正在补齐，可先查看下方英文原文。

Following up on my Qwen 3.6 port, I wanted to keep adding models and ended up fixing a bunch of things along the way, so it's its own engine now: Mference. Same core idea from TurboFieldfare, MoE models activate a few B params per token, so keep the shared core and KV cache resident and stream