两台 DGX Spark 跑 M3.1 Flash：代码 29.4 tok/s，约为 DeepSeek 三分之一
TP=2 本地实测：M3.1 Flash 代码 29.4 tok/s、工具调用 27.8 tok/s、131K 上下文；同配置 DeepSeek 约 90 tok/s。原因是每 token 读约 7.7GB（DeepSeek 约 3.8GB），Spark 解码受带宽限制。
This one feels a bit disappointing, honestly. @MiniMax_AI MiniMax-M3.1-Flash on two DGX Sparks ALPHA: 29.4 tok/s on code 27.8 on tool calls 131K context. My DeepSeek recipe does ~90 on the same TP=2 setup. MiniMax reads about 2x the bytes per token, and Spark decode is bandwidth-bound. Built on @MiaAI_lab's MiniMax-M3.1 TensorFold patch and @ashxhart TensorFold. Model and DSpark drafter by MiniMax. Decoding on a Spark is bandwidth-bound, so speed comes down to bytes read per token. MiniMax has ~24.7B active params, 64 attention heads at 8,192 wide and a big shared expert, all running on every token. ~7.7 GB per Spark per token. DeepSeek Flash is ~16B active and ~3.8 GB. I suggest a 2100-2200 MHz GPU clock cap, as I noticed it approaching heat cutoff at times. I am working on a performance pass right now and more cache work, but at the current speed and context, it's hard to justify using it. Early indicators show a potential 30% increase, but ultimately a smaller quant will be needed to get any speed. Recipe and weights in the reply. What would you run on it?