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

EnactraAI reports MiniMax M3.1 Flash at 73.3 overall on BuildingBench, ranking #12 and beating SWE-2's 67.3 while nearing DeepSeek V4.1 Flash's 72.9, with a best reconstruction score of 88.8.

MiniMax says the anonymous early build topped OpenCode for 12 straight days, and the finished M3.1 Flash is now live and free on OpenCode for three days.

The author tested M3.1 Flash on Keygen Bench and found it disappointing, calling it a case of "bad small model smell," with a video attached.

A video pits MiniMax M3.1 Flash against DeepSeek V4.1 Flash, but the post gives no task details or results, so you have to watch.

Official details for MiniMax M3.1 Flash: multimodal coding model with 1M context and tunable thinking, 62.2 on Automation-Bench, 91.0 on BabyVision, 73.7 vs DeepSeek's 77.0 on DeepSWE; API live now, weights October 23.

VentureBeat reports MiniMax M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark at lower cost, but carries a high hallucination rate.

CashuPayServer ran AI security audits across models; MiniMax M3.1 (Space Bunny) alone found a higher-severity bug all prior scans missed, and the author urges using it free on OpenCode.

Both models recreated the same viral Tesla Roadster police chase scene; MiniMax M3.1 Flash cost $2.34 and DeepSeek V4.1 Flash cost $3.07.

Compares per-1M-token prices (Opus 5.5 $4/$20, DeepSeek Flash $0.30/$1.20, M3.1 Flash $0.15/$0.60) and official scores: Automation Bench 62.2 beats Opus 5, Terminal Bench 38.1 lags; Space Bunny topped OpenRouter for 10 days on 56.7T tokens.

The author claims M3.1 Flash beats GPT-5.6 Sol, that Opus 5 is slightly better but 40x pricier, and that the model is highly efficient at 200 t/s.

Running the same 3D prompts from an Oct 6 challenge one-shot with no fixes, the Eiffel Tower came out worse than M3 while the fish and cat were solid; three scenes cost $0.19 versus $13.83 for Claude Opus 5.5.

The author ran 10 full bull/bear debates on the same two trade ideas with each model, no fallback, to compare how they perform in their trading-debate workflow.