Space Bunny (M3.1 Flash stealth) 3D test: 400–700 tok/s on the same koi pond prompt
The author reran the same koi courtyard 3D prompt on M3.1 Flash in MiniMax Code, getting far more detail than older MiniMax models at 400–700 tok/s versus roughly 250–300 tok/s for DeepSeek V4.1 Flash, using 88.94M tokens at no extra cost during the free period.
holy shit MiniMax M3.1 Flash is really cracked at 3D 💀 remember Space Bunny Alpha? it was getting crazy attention for its 3D demos during stealth. that was actually an early version of M3.1 Flash. the final release is even better, and MiniMax says it's nearly 3x faster at inference than its previous generation. so I gave it the same Koi Courtyard test I gave DeepSeek V4.1 Flash 1017, this time using MiniMax Code. and fuck, the result is incredible. the depth of the pond, the way you can see through the water, the rocks, swimming koi, bamboo fountain... everything looks SO good. compared to older MiniMax models, this feels like a MASSIVE leap in 3D. I'm genuinely amazed a Flash model can pull off this level of detail now. and the SPEED is another thing. I was seeing 400–700 tokens/sec at times, compared to roughly 250–300 tokens/sec for DeepSeek V4.1 Flash 1017 on faster setups. total usage: 88.94 MILLION tokens cache reads: 86.16 MILLION tokens and since I ran this on my MiniMax M Plan, it cost me $0 extra. didn't even burn my limits because M3.1 Flash usage is unlimited until Oct 25. they're also offering 50% off the first month of M Plan until Oct 31. 89 MILLION tokens for a fucking fish pond AGAIN 💀