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

A user reports that running a skill on free MiniMax M3.1 surfaced critical bugs that Opus, Fable and Astra/Sol did not find.

Given a Node project with three bugs, M3.1 Flash spent 15 minutes analyzing without edits on the broad first pass; after narrowing scope it changed the code and Codex independently ran the original tests at 5/5.

In 20 minutes at no extra cost, MiniMax M3.1 Flash produced a single-file web OS with boot screen, window manager, terminal, files, paint, snake, a Three.js galaxy and a built-in AI assistant.

A user remarks on MiniMax's long free periods — two weeks as Space Bunny and three more days for M3.1 Flash — and says the model's results are a pleasant surprise.

The author used a single prompt to have MiniMax M3.1 Flash build a playable Tokyo street endless runner with 20+ characters, weekly leaderboards and sound effects, citing low token cost.

A columnist notes MiniMax cut M3.1 Flash API prices by more than half, the stock traded below its IPO price that morning, and says the real test is the upcoming M3 Pro.

Same prompt test: M3.1 Flash produced a cinematic scene with fog beams, in-cloud lightning and sea foam, while DeepSeek V4.1 Flash just built a lighthouse; both start at $0.15/$0.60 per million tokens but DeepSeek doubles at weekday peak hours.

On the same 3D plush spider challenge, the author says M3.1 Flash stood out for concept-to-3D translation and plush texture, and lists 200+ tok/s, agentic workflows and 70% lower cost.

A roundup says M3.1 Flash is the official Space Bunny, citing self-reported ALE 29.6%, 200+ TPS, and ¥1/¥4 per million tokens, while noting several benchmarks are unverified and weights are closed.

A solo developer spent 5 hours with M3.1 Flash to build an interactive Three.js ring ship with part disassembly and docking, crediting strict layered acceptance criteria and cheap long-context retries.

A single prompt had M3.1 Flash compute keyframes in code for a line-art animation and synthesize water-drop and white-noise audio with numpy in one task, no external assets, costing under $0.05 at $0.15/$0.60 per million tokens with 1M context; the author notes it is OpenRouter's Space Bunny.

A short tip that vLLM is the way to serve MiniMax M3.1 Flash yourself, without memory, quantization or throughput details.