MiniMax-M3.1-Flash

How it is built

MiniMax-M3.1-Flash is a native multimodal mixture of experts. Its card gives the size as approximately 428B parameters, with 23B activated for each token.¹¹1MiniMax, MiniMax-M3.1-Flash model card on Hugging Face: "Approximately 428B total; 23B activated per token"; context window up to 1,000,000 tokens; three changes against M3 (sparse attention throughout, quantization-aware training, DSpark speculative decoding). https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash

The language model has 60 layers. The first 3 are dense, and the other 57 are expert layers.²²2MiniMax-M3.1-Flash config.json, read 23 October 2026: num_hidden_layers 60, moe_layer_freq 0 for the first 3 layers and 1 for the other 57, num_local_experts 128, n_shared_experts 1, num_experts_per_tok 4, num_attention_heads 64, num_key_value_heads 4, sparse_attention_freq 1 in all 60 layers, vision_config num_hidden_layers 32, quantization_config mxfp8 with moe_quant_algo NVFP4. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/config.json. Arithmetic: 60 minus 3 = 57. Each expert layer holds 128 routed experts and 1 shared expert, and the router sends a token to 4 of the routed ones.²²

Attention uses 64 query heads over 4 key-value heads.²² The architecture is the one of MiniMax-M3. A vision encoder of 32 layers handles images and video.²² Only text comes out.¹¹

What is new

M3.1-Flash makes three changes to M3.¹¹ Sparse attention now runs in all 60 layers, where M3 kept the first 3 in full attention.¹¹²² The routed experts and the attention paths are quantized during training.¹¹ A speculative decoding method called DSpark replaces the earlier multi-token prediction head, and its draft head ships with the weights.¹¹

MiniMax says online inference is nearly three times faster than before and that the usage cost of agentic tasks falls by 70%.³³3MiniMax, "MiniMax M3.1 Flash: Redefining Agent Performance and Cost", 19 October 2026: nearly three times faster online inference, usage cost 70% lower for agentic tasks, DeepSWE v1.1 73.7%, CyberGym 82%, flat pricing of $0.15 input, $0.60 output and $0.015 cached input per million tokens, available through the MiniMax API. https://www.minimax.io/blog/minimax-m31f It reports 73.7% on DeepSWE v1.1 and 82% on CyberGym, measured by MiniMax.³³

Using it

The model opened on the MiniMax API on 19 October 2026, with a context window of 1,000,000 tokens.³³⁴⁴4MiniMax, text generation guide, read 23 October 2026: MiniMax-M3.1-Flash context window 1,000,000; effort accepts low, medium, high, xhigh and max, and omitting it defaults to max. https://platform.minimax.io/docs/guides/text-generation Effort takes five levels, from low to max. When it is omitted, the default is max.⁴⁴

Pricing is flat, with no second tier for long prompts. A million input tokens cost $0.15, or $0.015 from the cache, and a million output tokens cost $0.60.³³

The weights are on Hugging Face and need no access request. The repository holds 72 files that total 252.4 GB.⁵⁵5Hugging Face model API listing for MiniMaxAI/MiniMax-M3.1-Flash, read 23 October 2026: 72 files, summed file sizes 252.4 GB. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/tree/main The config marks the experts as MXFP8 with NVFP4 packing.²²

Non-commercial use is free under the MiniMax Community License. A commercial product must carry a visible "Built with MiniMax M3.1-Flash" credit, and its maker must send MiniMax a one-time notice by email. Above US$20 million of yearly revenue, written authorization comes first. Military use is prohibited.⁶⁶6MiniMax Community License of MiniMax-M3.1-Flash, clauses 2 and 3 and the appendix of prohibited uses. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE