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

Same 'build what you think you look like in Minecraft' prompt run on MiniMax M3.1 Flash and DeepSeek V4.1 Flash, with estimated costs of $0.3 vs $0.38 and a side-by-side video.

The author used a single MiniMax M3.1 prompt to build an interactive 3D simulation of nature reclaiming an abandoned city, noting some parts are still rough.

The author agrees DeepSeek V4.1 Flash is better than MiniMax M3.1 Flash for the most part, but wants to give M3.1 Flash a fair shot inside MiniMax Code (MiniMaxAgent) and report back.

SAGE EXL3 2.15bpw quantization with TP2 over RoCE on two DGX Sparks gives 300K context, 726K KV pool, up to 39 tok/s single stream and 65–78 tok/s at four users.

The author ran five 3D frontend cases — retro CRT game, pixel sky castle, W16 engine, record wall, dune erosion — across four models and asks viewers which did best.

The promised open weights arrived: 428B total parameters, 23B active per token, 1M context, text/image/video input, with day-0 vLLM support for Blackwell, Rubin and MI350.

A user who spent about a month building small games with Space Bunny says it turned out to be MiniMax M3.1 Flash and was more capable than initially assumed.

The author made a full 57-second ad with MiniMax Code and reports scores of 240 for Space Bunny and 246 for M3.1 Flash, with all 52 checks passing on one agent task.

A user ran MiniMax M3.1 Flash locally on four DGX Sparks with vLLM and DSpark speculative decoding, handling six ~100K-token prompts at 84 tok/s combined decode and ~1,480 tok/s prefill.

Hands-on test of M3.1 Flash (formerly Space Bunny): coding roughly matches DeepSeek V4.1 Flash-1017, enterprise workflows and computer use lead domestic models, and 300 tasks cost about one yuan.

On a shared SVG prompt, Kimi K3.1 Pro Max produced more detail than MiniMax M3.1 Flash but drew the plane wrong, and the author still rates it stronger overall.

A user calls M3.1 Flash cheap and useful enough to substitute for DeepSeek, switching to it whenever DeepSeek rate limits stall.