MiniMax-M3.1-Flash Weights Arrive, and Four Sparks Run Them on Day One
428B total, 23B active, about 252 GB; a four-Spark recipe hits 46.5 tok/s on prose.
Released Thursday evening, running on four Sparks within hours.
MiniMax-M3.1-Flash open weights are live, @MiaAI_lab posted: 428B parameters, 23B active, weights about 252 GB. The model card adds a context of up to 1M tokens, text, image and video inputs, sparse attention in every layer, quantization-aware training and a DSpark draft head for speculative decoding.
@NVIDIAAI reposted NVIDIA’s day-zero support, with NVFP4 checkpoints on Hugging Face. Its figures, up to 200K TPS per GPU and more than 700 TPS per user, are for Vera Rubin, not Sparks.
On four DGX Sparks, @majewskizby posted the first measured bring-up on MiniMax’s own NVFP4 experts and MXFP8 weights with the bundled DSpark draft. His first numbers: 39.7 tok/s prose and 62.2 structured at one stream, and 166 tok/s prose in aggregate at 16 streams, with 262k context.
| Test | Result |
|---|---|
| speedbench, 73 prompts | 45.5 tok/s geometric mean |
| sparkDash prose, one stream | 46.5 tok/s |
| sparkDash structured, one stream | 85.0 tok/s |
| Cold prefill, 4K to 128K | 1,421 to 1,490 tok/s |
| qeval, 75 tasks | 73 of 75 |
| Needle retrieval | Pass at 116K and 218K tokens |
GB10 is sm_121, and some of MiniMax’s defaults are SM100-only, the repo notes. The NVFP4 KV cache does not run, so the recipe uses FP8 KV, Triton sparse attention and Marlin expert kernels. About 63 GB of weights sit on each box.
The license is not MIT, @superalesha notes. He reads it as free for commercial use with a “Built with MiniMax” credit, with permission needed above $20M in yearly revenue.
Sources
- @MiaAI_lab on X · Oct 22, 10:52 PM PT
- @MiaAI_lab on X · Oct 22, 10:52 PM PT
- @NVIDIAAI on X · Oct 22, 7:09 PM PT
- @majewskizby on X · Oct 23, 12:03 AM PT
- @superalesha on X · Oct 22, 11:47 PM PT
- MiniMax-M3.1-Flash model card · huggingface.co
- knapcio/MiniMax-M3.1-Flash-4x-DGX-Spark-TP4 · github.com