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

Three single-sentence prompts in MiniMax Code produced voxel and scene results in 1h48m, 1h50m and 1h6m with no extra prompting, plus DeepSWE 73.7% and API pricing.

First measured bring-up of M3.1 Flash with TP4 across four DGX Sparks: 39.7 tok/s prose and 62.2 tok/s structured decode at c1, 166/261 tok/s aggregate at c16, 262k context with ~1.37M-token FP8 KV cache, using MiniMax's NVFP4 experts and MXFP8 weights, with a full recipe posted.

Breakdown of the released weights: 428B total and 23B active params, 1M context, text/image/video input, sparse attention in every layer, ~250GB thanks to train-time quantization, a bundled speculative draft head, day-0 vLLM/SGLang support, and 62.8 vs 60.5 over DeepSeek V4.1 Flash on Automation-Bench; the license is free with credit but needs permission above $20M revenue.

Given a single-HTML three.js koi pond prompt, MiniMax M3.1 Flash one-shotted it in 11 minutes using 55k output tokens for $0.03, with 12 koi, a waterfall, lily pads and fireflies that swim over when fed.

Roundup of the official release: Space Bunny was an earlier M3.1 Flash, 428B total/23B active with text, image and video input and 1M context, weights on HF/ModelScope, plus scores like DeepSWE 73.7% and OSWorld V2 61.7 and day-0 support from Together and SGLang.

Input drops to ¥1 and output to ¥4 per million tokens, while Artificial Analysis measures 43 points — above Zhipu, Alibaba and Xiaomi Flash-tier models but below DeepSeek V4.1-Flash-1017; the author notes subscription quotas are shrinking at the same time.

MiniMax released M3.1 Flash weights: 428B parameters with 23B active per token, reporting 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2, and confirming the anonymous early version was Space Bunny.

Weights land on Hugging Face and ModelScope: 428B total/23B active MoE, DeepSWE 73.7 and OSWorld V2 61.7, with API pricing at $0.15 input and $0.60 output.

Open weights under the MiniMax Community License with 428B total and 23B active parameters, 1M context, 100T tokens in 72 hours, day-one support from ten chip vendors and six clouds, priced at $0.15/$0.60 per 1M tokens.

A Chinese rundown of the official M3.1 Flash specs: native multimodal MoE with 428B total and 23B active parameters, 1M-token context, text/image/video input, benchmark scores and the Hugging Face weights link.

Weights are on Hugging Face for self-deployment under a community license that requires attribution and notice for commercial use, plus a separate license above a revenue threshold.

An SGLang-Omni member analyzes the 428B MoE model's sparse attention, quantization-aware training and speculative decoding, citing OSWorld 61.7 and DeepSWE 73.7%.