In case you're not aware already, the DeepSeek V4 series is finally getting supported on llama.cpp with this PR ! The PR is at a very early stage right now, so only try it if you're consciously willing to experiment out of curiosity and accept severe stability/performance tradeoffs. It runs very slow (5-6 tps), GPU and FA support need work, etc., but it is reliable-enough already for correctness. This is my most anticipated model and I had some time to spare, so I ended up downloading the H...