Builders got MiniMax M3.1 Flash's open weights from 252.4 GB down to 93 GB in 10 days, and the lab thanked them by name

MiniMax posted a thank-you yesterday for everyone quantizing M3.1 Flash (428B, 252.4 GB), and I went through the ones they named. Numbers are each builder's, separate eval sets, not comparable. 128 GB PC: Maya-XS by PeasantSmith, 93 GB. 88.8% top-1 agreement (tool calls, agents). One DGX Spark: SAGE-EXL3 by ViC305, 108.64 GiB. 96.89% top-1, held-out sequences. 24.37 to 29.72 tok/s decode at 48K context. 160 GB+: Unsloth UD-Q2_K_XL , 143.9 GB. 91.5% top-1 weighted, seven categories. 4x 3090: superalesha , own engine, Unsloth 2-bit, 23 to 31 tok/s decode, 1120 tok/s prefill at 16K. Also: oMLX by jundotkim/mlx-community (Apple Silicon, oQ3e 196.2 GB), AMD Quark MXFP4 for MI350/MI355, Mia-AiLab EXL3-EAQ 3.55 bpw, Atomic Chat GGUFs. It helps that the model was trained with quantization baked in, sparse attention in every layer, 23B of 428B active per token. It's a coding and agent model that takes image and video input, up to 1M context if your box has the memory. People don't put this calibration work into weights they won't keep running, and superalesha ran out of RAM at 128 GB, went to 512 to keep going. Only used it hosted myself, agent jobs got done this morning, so I get why. Anyone running one of these on real coding or tool-call work?