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

A user tests MiniMax M3.1 Flash on 3D generation, saying it performs far better than expected and the pricing is extremely low, with a video demo.

On the same prompt, DeepSeek V4.1 Flash took 1h28m, 89.5M tokens and $0.54, while MiniMax M3.1 Flash took 1h6m, 8.12M tokens and $0.22.

A user notes that every session, question and answer for the M3.1 Preview 'Space Bunny' on OpenRouter passed through MiniMax's servers, so all that data was captured regardless.

A short relay of the official launch points: MiniMax M3.1 Flash claims nearly 3x faster inference and an API price cut of more than 50%, with no independent testing.

The author confirms M3.1 Flash is Space Bunny, says its computer use leads open models by a wide margin, it is fast on complex tasks with a non-robotic chat style, at $0.15/$0.60 per 1M tokens.

On the Extended NYT Connections benchmark, V4.1 Flash 1017 (high) rises from 89.6 to 92.8, while MiniMax M3.1 Flash (high) scores just 28.3.

Building an FPS game with MiniMax M3.1 Flash and minicode produced crude visuals, no vertical camera movement and a poor gun model, which the author calls shocking.

One run each on the same three.js scene: Claude Fable 5.1 took 10:21 for $2.82, GPT-6 Astra 1:44 for $0.53, while M3.1 Flash hit the 30-minute cap still refining at $0.18 without finishing.

Roundup of MiniMax M3.1 Flash's official launch details: Space Bunny was its early build, 428B MoE with 23B active, 1M context, $0.15/$0.60 per million tokens, ~3x faster inference, open weights on Oct 23.

MiniMax M3.1 Flash cuts input and output prices by over 50% and claims a 70% drop in agentic task cost; the post says it is roughly level with DeepSeek V4.1 Flash 1017 on DeepSWE in both performance and price.

Notes that MiniMax M3.1-Flash arrived two days after V4.1-Flash and promises six points on choosing between the two domestic agent models and lowering batch-task costs.

A blogger says MiniMax M3.1 Flash answers benchmark questions instantly but cannot name its own boss, suspecting the model memorized test sets rather than showing real ability.