MiniMax M3.1 Flash in practice: tests and comparisons with DeepSeek V4.1 Flash

MiniMax officially released M3.1 Flash open weights, with Together Compute as a day-0 launch partner serving the model in production on its AI Native Cloud.

A DFlash 2 speculative decoding drafter trained for MiniMax M3.1 Flash is released under Apache 2.0, making generation up to 3.5x faster while remaining lossless.

inco_ai released an Apache 2.0 DFlash 2 drafter for MiniMax M3.1 Flash that speeds up generation up to 3.5x without loss, with a Hugging Face link.

A user breaks down the open weights: 428B total and 23B active parameters, 252GB official weights, 239.5GB AMD MXFP4 quant, tight fit on two DGX Sparks, and ¥1/¥4 pricing matching DeepSeek off-peak.

A user suspects M3.1 Flash is just M3 rebranded, with attention sparsified in three more layers, quantization and DSpark added on top.

A user summarizes that M3.1 Flash can be self-hosted with 428B total and 23B active parameters and 73.7% on DeepSWE, advising multi-GPU owners to download weights and compare against API costs.

A user notes that MiniMax has published the MiniMax-M3.1-Flash weights on Hugging Face, linking to the model repo.

M3.1 Flash weights are on Hugging Face: 428B total / 23B active parameters, 1M context, self-reported BankerToolBench 76.2 matching Opus 5 but lower OSWorld 2.0 and Terminal Bench 4.0; commercial use needs attribution and written permission above $20M revenue.

The poster spots the MiniMaxAI/MiniMax-M3.1-Flash weights repo on Hugging Face, meaning the model can be downloaded and self-hosted.

A rundown of the newly released M3.1 Flash weights: a 428B-total, 23B-active MoE with sparse attention and DSpark speculative decoding, official scores of 73.7% on DeepSWE v1.1 and 82.2 on CyberGym, with community quants and vLLM throughput still to come.

MiniMax released open weights for M3.1 Flash: 428B total parameters, 23B active, 1M context, and text/image/video input, confirming Space Bunny was its early version.

A user relays M3.1 Flash open-weight specs — 428B total parameters, 23B active — along with a DeepSWE score of 73.7%.