MiniMax M3.1 Flash 428B (23B active)
4× NVIDIA RTX 3090 · 16,384 ctx
- generation:
- 23-31 tokens/s
- quant:
- 2-bit (GGUF)
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
User rounds up community quantizations of MiniMax M3.1 Flash (428B total, 23B active), whose open weights are 252.4 GB, and reports each builder's own figures. The one local run with numbers is superalesha's on 4x RTX 3090 with an own engine and Unsloth 2-bit weights: 23 to 31 tok/s decode and 1120 tok/s prefill at 16K context. The builder ran out of RAM at 128 GB and moved to 512 GB to continue. Other builds listed: Maya-XS at 93 GB for a 128 GB PC, SAGE-EXL3 at 108.64 GiB on one DGX Spark with 24.37 to 29.72 tok/s decode at 48K context, Unsloth UD-Q2_K_XL at 143.9 GB, plus oMLX, AMD Quark MXFP4, EXL3-EAQ and Atomic Chat releases. The user says they have only used the model hosted.