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

MiniMax Code's R&D lead announced M3.1 Flash, free and unmetered for subscribers until October 25 with no credit consumption.

The author recaps MiniMax claiming its stealth Space Bunny as M3.1 Flash and cites community tests: steady 85–112 tokens/s, a complex 3D scene plus code built in under three minutes, near-instant multi-file fixes.

Spec roundup: Space Bunny was M3.1 Flash all along, with 1M context, adjustable thinking depth, $0.165 input / $0.66 output per 1M tokens, $0.0165 cache, and open weights coming soon.

A user notes that whether MiniMax M3.1 Flash fits on a single DGX Spark will decide its local matchup against DeepSeek V4.1 Flash, which runs across two machines; no parameter count is known yet.

A developer calls the API with max thinking and generates a Pagoda floating-island scene in a single 50K-token pass with no multi-turn or tools, rating it above DeepSeek V4.1 Flash on similar work.

A user sums up M3.1 Flash specs: 1M context, image and video input, adjustable thinking effort, and $0.165 input / $0.66 output per 1M tokens with $0.0165 cache reads.

The author says Space Bunny felt extremely fast in testing and produced better results than Opus 5 and GPT 6 Sol, suggesting M3.1 Flash may combine speed, efficiency and Opus 5-level quality.

A roundup of MiniMax's teaser and known specs: multimodal coding model, 1M-token context, five thinking effort levels, OpenAI/Anthropic-compatible APIs; early testers call it fast and cheap, while one found DeepSeek V4.1 Flash made fewer errors.

Testing via an M Plan subscription, the user gets ~200 tok/s from M3.1 Flash; both models finish the same test in 38 minutes, but MiniMax's output quality is preferred, and the run used only ~6% of the 5-hour limit on the $22 Go plan.

A video test shows a pagoda generated in one shot with no harness at 50K tokens, which the author rates above DeepSeek V4.1 Flash.

Space Bunny has come out of stealth as MiniMax's new M3.1 Flash model, with nothing known yet beyond its 1M-token context window.

A timeline of the official tease: Space Bunny went anonymous and free on OpenRouter and OpenCode on Sep 23, was identified as MiniMax via tokenizer fingerprints, and the API docs now drop the Preview suffix.