MiniMax releases MiniMax-M3.1-Flash

Background

M3.1-Flash came after M3 in the same line of models. The new version changes the attention design and the decoding method, and the model has about 428 billion total parameters ^[1].

What happened

MiniMax posted the weights of M3.1-Flash to Hugging Face on 17 October 2026 ^[2]. Hugging Face is a platform where researchers and developers share models and datasets, so the weights became available for download there.

Key details

The model has about 428 billion total parameters, with 23 billion activated, up to 1 million tokens of context, and text, image and video input ^[1]. It uses MiniMax Sparse Attention throughout, replacing the first three full-attention layers that M3 kept ^[1]. Quantization-aware training covers the routed experts and attention paths to cut memory use ^[1].

DSpark speculative decoding replaces the EAGLE-like multi-token prediction used in M3 ^[1]. Speculative decoding lets a fast component propose several tokens that the main model checks, which can speed up generation. Attention and decoding both affect speed and memory use, which is why changes to them are often reported in terms of running costs.

On benchmarks, MiniMax reported 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 82.2 on CyberGym and 76.2 on BankerToolBench ^[1].

Terms used here

Speculative decoding is a generation technique in which a faster component drafts candidate tokens and the main model checks them, so the final output is the one the main model accepts. Quantization stores a model's numbers at lower precision, which reduces the memory needed to hold it. Quantization-aware training teaches the model to work well at that lower precision.

Significance

The release is mainly a change in how the M3 line runs: the attention layers, memory use and decoding method are the parts MiniMax changed, and the figures above are its own reported results ^[1].