MiniMax-M3.1-Flash

About

MiniMax-M3.1-Flash (MiniMaxAI/MiniMax-M3.1-Flash) is a native multimodal mixture-of-experts (MoE) model for coding, agentic workflows, computer use, and cybersecurity, released by MiniMax on 2026-10-19 (official announcement blog "Redefining Agent Performance and Cost"; Space Bunny, which processed 200+ trillion tokens while operating anonymously, was an early version of M3.1-Flash). It retains the MiniMax-M3 backbone: ~428B total parameters with ~23B activated per token, a 60-layer transformer with 128 routed experts (4 active per token plus 1 shared expert, sigmoid routing scoring), grouped-query attention with 64 attention heads / 4 KV heads and head_dim 128, and a context window of up to 1,000,000 tokens (config max_position_embeddings = 1,048,576). The model accepts text, image, and video input and produces text; vision is handled by a CLIP-style vision tower (32 layers, patch size 14, 3D-RoPE, dynamic-resolution image tokens with patch-merge token compression).

M3.1-Flash makes three changes to M3's core architecture. First, MiniMax Sparse Attention (MSA) is used throughout the model - the config carries a sparse-attention overlay (sparse_topk_blocks=16, block_size=128, index heads=4) on the GQA baseline, and the first three full-attention layers of M3 are replaced with sparse attention layers. Second, quantization-aware training (QAT): the routed MoE experts and attention paths are quantized during training under an mxfp8 dynamic-activation scheme with NVFP4 pack-quantized expert weights, reducing memory footprint and improving serving efficiency (the HF Hub lists "220B params" for the shipped NVFP4-packed weights, while the config.json-derived full parameter count is ~428B as on M3). Third, DSpark speculative decoding replaces the earlier EAGLE-like MTP approach; the release ships the DSpark draft head for use with compatible serving implementations (SGLang, vLLM, TokenSpeed, Transformers integrations in progress; multi-chip deployment via FlagOS).

On agentic benchmarks the card reports 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 29.6 on Agents' Last Exam, and 82.2 on CyberGym. Recommended sampling settings are temperature=1.0 and top_p=0.95; reasoning_effort is configurable in the chat template (low, medium, high, xhigh, max). MiniMax prices M3.1-Flash at $0.15 input / $0.60 output / $0.015 cached input per 1M tokens, claims online inference is nearly 3x faster with 170-200+ TPS output, and says overall agentic usage cost is 70% lower than its predecessor. Under the MiniMax Community License, weights are openly released on Hugging Face (34 third-party quantizations, 9 finetunes).

Training Data Quantization-aware training: routed MoE experts and attention paths are quantized during training (mxfp8 dynamic-activation activations, NVFP4 pack-quantized expert weights).