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

A single prompt kept MiniMax M3.1 Flash building for 2 hours 27 minutes straight, producing a voxel Japanese garden with pagoda, koi pond, cherry trees and a dragon, with every prompt detail present.

A user posts a comparison image and says GLM 5.3 Flash can no longer match MiniMax M3.1 Flash, urging GLM to ship its next model.

A motion-design animation done entirely on MiniMax Cloud used about 80M tokens on M3.1 Flash, roughly $0.29 under the token plan, while DeepSeek used about 68M tokens at $0.53–1.06; the author finds quality close but Flash much cheaper for bulk design work.

Hands-on test of MiniMax M3.1 Flash generating a 3D world scene, with the whole autumn valley project costing about $1.62; the author calls it the budget king versus Argon, DeepSeek V4.1 Flash and Opus 5.5.

The author recreated MiniMax's promo video in p5.js with M3.1 Flash inside MiniMax Agent, adding retro film grain, and praises its speed-quality-cost balance plus shot composition and visual storytelling.

One prompt produced Halo, a dashboard with a live 3D globe tracking sales, animated revenue charts and real-time transactions, showing flash models handling complex UIs.

Using a single prompt and an open-source skill, the author had M3.1 Flash produce a presenter-style explainer video for the DeepInfer project, was happy with the result, and notes it is currently free so Opus 5.5 isn't needed.

Cites independent benchmark data: M3.1 Flash scores 42, 13 points above M3, at about $0.22 per task, but hallucinates on roughly 93% of fact questions.

A single prompt produced a space mission control dashboard with an interactive 3D Earth, satellite tracking and orbits, telemetry panels and glass UI; the author was surprised by the visual complexity and found it fast and cheap on tokens.

Questflow's own benchmark puts MiniMax M3.1 Flash at 94.9, up 21 points from M3's 73.8 and just 2.8 behind Claude Opus 5.5 at 97.7.

The author likes M3.1 Flash's autonomy but finds total spend higher, preferring DeepSeek V4.1 Flash's cheaper cache reads.

Breaks down M3.1 Flash's $0.15/$0.6 pricing, the Space Bunny reveal and Oct 23 open weights, citing Artificial Analysis's 120 tok/s and $0.22 per task (57% drop) to question the official 70% claim.