MiniMax confirms 'Space Bunny' was an earlier M3.1 Flash release

The anonymous model gets a name

MiniMax has confirmed that "Space Bunny", the model it released without a public identity, is an earlier version of M3.1 Flash. According to a dev.to write-up, the identification came in a 19 October announcement from the company, settling a launch that had circulated under a codename rather than a product name.

The announcement also draws a line between two things that are easy to conflate: access to a hosted model and access to its weights. Hosted routes are available now, while the weight files are promised for later. An official MiniMax post on X names 23 October as the intended date, which dev.to characterises as a release commitment rather than a completed download.

Hosted access is live, weights are not

As reported by dev.to, MiniMax lists three routes for using M3.1 Flash today: the API, the M Plan, and MiniMax Code. The company's model documentation names the model MiniMax-M3.1-Flash, with a one-million-token context window and an adjustable thinking depth.

dev.to cautions that a large context window is capacity, not quality: it says nothing about how well an agent selects the right evidence from a long repository. The adjustable reasoning setting is another variable worth recording when comparing runs.

Vercel is running a temporary free window

Vercel announced on 19 October that M3.1 Flash is free through its Gateway until 22 October when MiniMax serves the request, under the model identifier minimax/minimax-m3.1-flash.

Provider routing matters here. Citing Vercel's routing documentation, dev.to distinguishes between an order preference, which tries preferred providers first but can still fall back elsewhere, and a MiniMax-only allowlist, which makes an unavailable route fail rather than silently pick another provider. For a trial meant to stay on the promoted route, the allowlist is the safer configuration. Vercel says the pinned route returns to MiniMax's normal rates once the offer ends, so launch-week pricing should not be mistaken for the default.

Manufacturer numbers, not independent ones

MiniMax reports improved agent performance and inference speed. dev.to points out that these are manufacturer results, that its own write-up did not reproduce them, and that the earlier anonymous model should not automatically be treated as identical to today's production endpoint.

Before swapping the backend of an existing coding agent, dev.to suggests choosing tasks with an observable finish line — fixing a known failing test, making a small change covered by an existing acceptance test, and handling a tool error without silently abandoning the task — while keeping tools and acceptance criteria consistent across backends. A fast plausible answer and a completed repository change are different outcomes.

A checklist for evaluation

dev.to's practical recommendations include:

• Confirming which product or endpoint you are using and saving its exact model name.
• Recording the thinking setting and verifying the tool-call format your application receives.
• Judging task quality against the same acceptance criteria used for your current backend.
• Waiting for the actual weight files before planning local deployment, then inspecting their licence, serving instructions and hardware requirements.
• Keeping anonymous previews, production endpoints and downloadable checkpoints distinct in your notes.

For a hosted trial, the write-up also advises keeping a run record that ties together the provider's exact model identifier, the endpoint, reasoning settings, prompt, tool definitions and time — since a screenshot of a model name alone will explain little if behaviour changes later.

Why it matters

Anonymous launches generate attention but make results hard to reproduce or compare, because nobody can pin down what was actually run. Attaching a real product name to "Space Bunny" turns speculation into something testable. The gap between hosted access and downloadable weights is equally consequential: it determines whether a developer is evaluating someone else's backend or preparing to operate the model themselves. Temporary free promotions such as Vercel's add one more trap, since a cost figure observed during launch week says very little about what the same workload will cost afterwards.