DS4 lands DeepSeek V4 Flash on Apple Silicon

Hot take: this is more of a systems bet than a model runner, and that makes it interesting.

• –The project is optimized for one model, one hardware family, and one storage hierarchy, which is how it gets away with aggressive context handling.
• –SSD-backed KV cache is the real differentiator here; it reframes disk as a first-class extension of memory for long-context local inference.
• –The Metal-only choice keeps the implementation focused, but also makes the audience very specific: high-end Mac users who want local DeepSeek V4 Flash performance.
• –This is the kind of repo that matters if the model really does behave well under compression and long-context reuse, because the bottleneck shifts from raw model size to memory management.