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 the selected experts off SSD. What runs now: Gemma 4 26B-A4B — ~2 GB, 31–35 tok/s on a 24 GB M5 Pro Qwen 3.6 35B-A3B — ~1.45 GB, 19–23 tok/s DeepSeek-V4-Flash 284B-A13B — new. ~6.8 GB peak memory, mostly ~5.3 GB in practice, up to 4.8 tok/s on the same 24 GB M5. 2-bit dynamic quant, ~91 GB on disk. Also picked up a native Mac app with multi-turn chat, an OpenAI-compatible server, and local PDF/DOCX/PPTX/XLSX attachments along the way. From here I want to keep adding model families, cut the expert-read wait (decode is ~53% I/O right now, serialized with compute), and push context past 4K. Not very useful beyond a few turns but you can technically run a "usable" dsv4f on a 8gb Mac. It only gets better from here.