The anonymous model had a name
MiniMax has named M3.1 Flash as the model users had been testing anonymously as “Space Bunny”, and says that early version processed more than 200 trillion tokens before the reveal. The launch post puts online inference at nearly three times the speed of its predecessor and overall agentic usage cost 70% lower.
The throughput figures are specific: more than 200 tokens per second in everyday output, with a median speed consistently above 170. Pricing is flat at $0.15 per million input tokens, $0.60 per million output and $0.015 for cached input.
Six scores, all vendor-reported
Unusually for a Flash-tier release, MiniMax publishes numbers rather than adjectives. BankerToolBench: 76.2. DeepSWE v1.1: 73.7%. OSWorld V2: 61.7, which MiniMax says is first among Chinese models. Agents’ Last Exam: 29.6, also claimed first among Chinese models. CyberGym: 82%, which the company describes as leading on vulnerability discovery and validation. AutomationBench-AA gets a placement — “among the top performers” — but no score.
Every one of these is MiniMax’s own measurement. The post compares the model to unnamed “leading models” and “top models” without publishing a single competitor’s score, so the scores stand as claims about MiniMax’s own system rather than as a ranking anyone can check.
What is missing
The post gives no parameter count and no context window — two numbers a developer needs before committing to a model, and two that MiniMax has disclosed for previous releases. It says the weights will be released later, and that a more capable Pro version will follow, without dates for either.
Availability is immediate through the MiniMax API, the M Plan subscription and MiniMax Code.
The Space Bunny disclosure is the interesting part
Shipping a model anonymously into a third-party coding tool and collecting 200 trillion tokens of real usage before announcing it is a different kind of launch from publishing a benchmark table. It gives the vendor evaluation data no leaderboard produces, and it gives users no way to know which model answered them.
That practice is becoming common enough to be worth naming. The question it raises is not whether M3.1 Flash is fast — the usage data presumably settles that internally — but whether a developer who spent weeks on “Space Bunny” would have chosen it had the name been on the tin.
