7.6
DeepSeek V4 Flash on a Single AMD MI300X — 168 tok/s for $1.99/hr, and What You Give Up
A community production stack runs DeepSeek V4 Flash (304B MoE) on a single AMD MI300X at 168.6 tok/s single-stream, 542 tok/s at 8 streams and 830 tok/s burst — no additional quantization, 256K context. Review of the patches, the FNUZ vs OCP FP8 trap, the $1.99/hr AMD Developer Cloud economics debate, and honest HN takes on whether self-hosting a model whose API costs pennies makes sense.