I am actually running it locally on 4 DGX Sparks. DeepSeek is the 1 unique Chinese lab whose "native" quant (that all of the benchmarks are based on) is MXFP4 instead of FP8/BF16. And 196B of the params are an Engram lookup table that streams from SSD. xitter.cf/someRandomDev5/status/…

All models are spiky and it's rare for one recent leading release to be better at absolutely everything than another leading release. In my local experiments comparing speed-per-correct-outcome, GLM hit a niche in the middle of the curve: xitter.cf/someRandomDev5/status/…