AI disclosure: This source-based explainer was researched and written with an AI agent. It is not a hands-on model benchmark. Sources checked on 20 October 2026.
Space Bunny has a name now. MiniMax’s 19 October announcement identifies the anonymous model as an earlier version of M3.1 Flash. The launch also separates two things that are easy to blur together: access to a hosted model and access to its weights.
The API, M Plan and MiniMax Code are available routes according to MiniMax. The same announcement says the weights will come later. Its official X post names 23 October as the intended date. That is a release commitment, not a completed download.
That distinction changes the first question a developer should ask. Are you evaluating a hosted backend, or preparing to operate the model yourself?
The free Gateway route has a provider condition
Vercel’s 19 October Gateway announcement says M3.1 Flash is free until 22 October when MiniMax serves the request. The Gateway model identifier is minimax/minimax-m3.1-flash.
Provider choice matters here. Vercel’s routing documentation distinguishes an order preference from an only allowlist. An order tries preferred providers first; it can still fall back elsewhere. For a trial intended to stay on the promoted route, a MiniMax-only allowlist makes an unavailable route fail rather than silently choose another provider.
Check the live endpoint prices and the provider recorded for the request. The promotion is temporary; do not treat a free rate observed during launch week as a permanent default. Vercel says the pinned route returns to MiniMax’s normal rates after the offer.
The name in your configuration matters
MiniMax’s model documentation lists MiniMax-M3.1-Flash, a one-million-token context window and adjustable thinking depth. Those are useful starting points for an integration checklist.
A large context window is capacity. It does not tell you how well a particular agent will select the right evidence from a long repository. An adjustable reasoning setting is another variable to record when comparing runs.
For a hosted trial, a useful run record would include the exact provider model identifier, endpoint, reasoning settings, prompt, tool definitions and time. Keep those together. If the run changes later, a screenshot of the model name alone will be a poor explanation of what happened.
A benchmark headline is a reason to investigate
MiniMax reports improved agent performance and inference speed. Those are manufacturer results. This article did not reproduce them, and the earlier anonymous model should not automatically be treated as identical to today’s production endpoint.
Before swapping a backend in an existing coding agent, choose a few tasks with an observable finish line. For example: fix a known failing test, make a small change with an existing acceptance test, and handle a tool error without silently abandoning the task.
Keep the tools and acceptance criteria consistent. Record whether the patch actually passes, how many attempts it takes and whether the agent needs intervention. A fast plausible answer and a completed repository change are different outcomes.
Those are suggested checks, not experiments claimed as completed here.
The practical checklist
- Hosted access: confirm which product or endpoint you are using and save its exact model name.
- Behavior: record the thinking setting and verify the tool-call format your application receives.
- Task quality: judge a concrete outcome with the same acceptance criteria used for your current backend.
- Local deployment: wait for the actual weight files, then inspect their licence, serving instructions and hardware requirements.
- Comparisons: distinguish an anonymous preview, a production endpoint and a downloadable checkpoint in your notes.
The reveal settles the identity question. The deployment question still depends on which route you need. As of this source check, MiniMax’s launch describes hosted access now and downloadable weights later. That is enough to build a sensible evaluation checklist; it is not evidence that a local installation or a workload migration has already succeeded.
Hands-on follow-up: Free MiniMax, zero API keys: build a read-only SQL report app.
Top comments (0)