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

Running DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3 and MiniMax M3.1 on two AMD Strix Halo boxes plus an RTX PRO 6000, the author traces most gains to config: GPU memory clock stuck in P1 costs ~4.5% bandwidth and 1.65x slower first tokens, and pipelined expert loading cut load time from 109s to 62s.

A short complaint says Kimi K3.1 Pro, MiniMax M3.1 Flash, Muse wonder and Gemini 4 Argon all flopped, leaving only DeepSeek V4.1 Flash 1017 in use, without any concrete failure details.

Given a small Python shop library with five planted bugs and one slow function, M3.1 Flash fixed all six with no hints, found a seventh, and cut a duplicate check from 23.8s to 0.01s, though it also changed unrequested behavior.

A single prompt had M3.1 Flash build a working 3D interior design app with drag-and-drop furniture, real-time materials and lighting, alongside claims of 200+ tok/s and 70% lower agentic costs.

A user posts a video comparing M3.1 Flash with DeepSeek, saying the gap is about to show, though the post itself gives no concrete results.

A user reports the same task consumed 30% of a 5-hour quota on GPT-6.1 Sol but exhausted three full 5-hour quotas on MiniMax M3.1 Flash.

On the same 3D Earth scene prompt, M3.1 Flash finished in 15 minutes for free during the promo while GPT-6.1 Sol took an hour and $3.20, with the author preferring M3.1 Flash's visuals and interactivity.

An OpenCode daily usage chart shows DeepSeek 4.1 Flash surging since 1017, MiniMax M3.1 Flash dropping into the others bucket after its free promo ended, and GLM 5.3 Flash a distant second.

On a TP=2 Spark setup M3.1 Flash gets 29.4 tok/s on code and 27.8 on tool calls at 131K context, versus roughly 90 for DeepSeek, because it reads about 7.7 GB per token against DeepSeek's 3.8 GB.

The author compares MiniMax M3.1 Flash at ¥1 per million input and ¥4 output tokens with DeepSeek's Flash update two days earlier, arguing the race has shifted from smartest to cheapest model that finishes the job.

M3.1 Flash weights are on Hugging Face and ModelScope: 428B total parameters with 23B activated per token, text/image/video input, 1M-token context, official 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0 under the MiniMax Community License.

A developer says DeepSeek V4.1 Flash produces vulnerable code in critical areas like authentication, GLM-5.3 finds flaws but writes insecure code, and Opus 5.5, GPT-6.1 Sol and MiniMax M3.1 Flash clearly do better there.