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

M3.1 Flash weights are downloadable: 428B total/23B active, 1M context, vLLM/SGLang/TokenSpeed support; free non-commercial use, attribution for commercial, written approval above $20M revenue.

EvoLink runs one 40-second Three.js motor animation prompt on both: DeepSeek V4.1 Flash 1017 took 2h14m for ~$0.83, MiniMax M3.1 Flash 2h43m for ~$0.25.

AtomicChat published GGUF quants of the 428B MiniMax M3.1 Flash on Hugging Face, shrinking it from 258GB to 203.5GB and recommending Q3_K_M for 256GB machines.

Using OpenRouter pricing ($0.1 in / $0.55 out per million tokens), a 2M-input, 200K-output agent task costs ~$0.31 on M3.1 Flash versus ~$12 on Opus 5.5, nearly 40x.

MiniMax M3.1 Flash, the model previously anonymous on OpenRouter as Space Bunny, fetched 503 arXiv papers, wrote 4,000+ lines of code, self-tested and fixed bugs in an 86-minute unattended run for $0.88; the same tokens would cost about $8.46 at Claude Sonnet 5.5 pricing.

Given only a creative brief, the agent produced a forest, lakeside abandoned house, French village and connecting paths as one scene; the author notes 70% lower agent cost and $0.15/$0.60 API pricing.

Advice on choosing between self-hosting the 428B/23B-active open weights (needs multi-GPU vLLM/SGLang, DeepSWE 73.7%) and the official API at ¥1 input / ¥4 output per million tokens, after estimating monthly token use.

The open-source edulab skills project now generates four styles of math solution videos with MiniMax-M3.1-Flash; the author reports surprisingly good results and cites 1 yuan/M input and 4 yuan/M output tokens.

A single motion-design prompt had M3.1 Flash build a brand promo video entirely in code, with transitions the author compares to Claude Opus 5.5.

MengTo made three keyboard sites, a SaaS page, a rainy 3D Singapore and a 3D lighthouse game with MiniMax M3.1 Flash, sharing the full prompt process and per-token pricing.

MiniMax released full M3.1 Flash weights on Hugging Face: a 428B-total, 23B-active MoE with 1M context and up to 520K output tokens per call.

Both models completed the same 3D gummy piano task, with MiniMax at 46 minutes and DeepSeek at 49; the author also compares API pricing, noting MiniMax is flat-rate while DeepSeek doubles at peak.