MiniMax M3.1: What’s Actually Official?

Yes—but the released model’s exact name is MiniMax-M3.1-Flash-Preview, not simply “MiniMax M3.1.” It is an official preview model that entered MiniMax Code on September 27, 2026, and opened to public beta on September 28; a final, generally available M3.1 release has not been confirmed in the available official materials.

Is MiniMax M3.1 an Official Model?

MiniMax M3.1 refers to the company’s M3.1 model project, while MiniMax-M3.1-Flash-Preview is the model ID currently available in MiniMax’s documentation and products. That distinction matters when searching for announcements, benchmarks, or API instructions.

The confirmed timeline is:

September 27, 2026: The preview became available inside MiniMax Code.
September 28, 2026: MiniMax announced the public beta, according to China Securities News.
October 17, 2026: The available evidence still identifies it as a preview rather than a final general-availability release.

MiniMax describes it as a production-oriented model for complete development tasks: locating problems, writing code, handling edge cases, adding regression tests, and checking the effects of changes. Some earlier community reports connected it with the anonymous “Space Bunny” model, but sources disagree about whether MiniMax directly confirmed that identity.

Confirmed M3.1 Flash Preview Specifications

The MiniMax API guide and reporting provide the following core details:

Feature	Confirmed detail
Official model ID	`MiniMax-M3.1-Flash-Preview`
Context window	1,000,000 tokens
Input types	Text, images, and video
Reasoning effort	`low`, `medium`, `high`, `xhigh`, and `max`
Reasoning behavior	Adaptive thinking is always required
Default effort	`max` if no effort setting is supplied
Main access	MiniMax Code and MiniMax subscription plans
Public weights	Not released
Official parameter count	Not published
Official model card	Not available in the reviewed sources

The 1M-token window makes the model suitable for large repositories, long documents, and extended agent sessions. Native multimodal input also allows a developer to combine code with screenshots, diagrams, design references, or screen recordings instead of first translating every visual detail into text.

One important API detail is that thinking cannot be disabled. A request that explicitly disables thinking or uses an unsupported none effort can return HTTP 400. Because the default is max, applications should set the effort explicitly; otherwise every call may consume more tokens and take longer than necessary.

What Has Been Confirmed About Performance?

MiniMax has not published a standardized M3.1 benchmark table, parameter count, speed guarantee, or model card. Therefore, benchmark claims attributed to M3.1 should not be treated as official specifications.

One independent KingBench 3 report, tested on September 28, recorded:

MiniMax M3.1 Flash Preview: 53 out of 80, or 66.25%
Claude Opus 5.5: 93.75% in the same run
M3.1’s strongest example: 9/10 on a folding-table 3D task
Weaker interactive examples: 4/10 for an archery game and 3/10 for an elevator simulation

This is useful evidence about strengths and failure modes, but it is not a universal ranking. A separate hands-on creative test reported that M3.1 produced competitive interface animations, pixel animation, and interactive visual projects against selected Opus 5.5 prompts. Those examples were subjective and did not constitute a standardized benchmark.

Claims seen elsewhere, including a 73.8% SWE-bench Verified result, a specific tokens-per-second figure, or a $0.10 per-million input-token price, could not be traced to a primary MiniMax source. Do not use those figures for procurement or API budgeting without further confirmation.

Pricing and Availability

M3.1 Flash Preview is not documented as an unrestricted, free model. Current access is connected to MiniMax Code and the company’s M Plan or Token Plan ecosystem. A Subscription Key is used with the documented API-compatible endpoints.

A secondary report documented unlimited M3.1 use for subscribers from October 1 through October 7, 2026 in MiniMax Code. That promotional period has ended, and it should not be interpreted as continuing free access.

There is also no confirmed standalone pay-as-you-go price for M3.1 in the reviewed sources. Do not confuse it with the stable MiniMax M3: LLM Reference lists M3 at $0.30 per 1 million input tokens and $1.20 per 1 million output tokens. That price belongs to M3 and should not automatically be applied to M3.1.

MiniMax M3.1 API Setup

The official documentation supports two protocol styles:

Protocol	Base URL	Key	Model ID
OpenAI-compatible	`https://api.minimax.io/v1`	Subscription Key	`MiniMax-M3.1-Flash-Preview`
Anthropic-compatible	`https://api.minimax.io/anthropic`	Subscription Key	`MiniMax-M3.1-Flash-Preview`

Use the OpenAI-compatible format for tools that accept custom OpenAI-style providers. Use the Anthropic-compatible format for coding clients built around that protocol.

For example, a client configuration should include:

Base URL: https://api.minimax.io/v1
API key: your MiniMax Subscription Key
Model: MiniMax-M3.1-Flash-Preview
Effort: low, medium, high, xhigh, or max

The existence of API-compatible endpoints does not necessarily mean M3.1 has open, standalone pay-as-you-go billing. Current documentation ties authentication to a MiniMax subscription key.

M3.1 vs MiniMax M3 and Competing Models

MiniMax M3 is the more established general model in the available catalog, while M3.1 Flash Preview is a newer, coding-focused preview with selectable reasoning effort. Both are associated with a 1M-token context and multimodal input, but their release status and access conditions differ.

Against competing models, the evidence is mixed:

M3.1 placed well below Opus 5.5 in the cited KingBench 3 run.
It performed competitively in selected creative-coding demonstrations.
It offers five explicit reasoning levels and a very large context window.
MiniMax has not supplied enough official benchmark data to claim category leadership.
Its principal advantage may be access through MiniMax’s plan and coding ecosystem rather than proven top-tier benchmark performance.

For a serious comparison, test task completion, test-pass rate, tool-call reliability, latency, total tokens, and cost on your own codebase. A large context window is useful only if the model retrieves and applies the relevant material correctly.

How to Test MiniMax M3.1

Open MiniMax Code on the web, macOS, or Windows and sign in.
Select MiniMax-M3.1-Flash-Preview.
Set low or medium for small edits and routine test generation.
Use high, xhigh, or max for repository-wide changes, difficult debugging, or architectural work.
Give the model a real task with acceptance criteria, test commands, and relevant screenshots.
Require it to inspect the repository, implement the change, run tests, and report every modified file.
Compare the result with your current model using the same prompt, context, tools, and time limit.
Record completion rate, test failures, duration, token usage, and any manual corrections.

Do not use MiniMax M3 benchmark results as a substitute for testing M3.1. The preview’s architecture, serving setup, or behavior may differ even when the model family and context size are related.

FAQ

Is MiniMax M3.1 an official model?

Yes. MiniMax-M3.1-Flash-Preview is an official MiniMax preview model. “MiniMax M3.1” is the broader name, while the longer name is the model ID used in products and API documentation.

When was MiniMax M3.1 released?

It appeared in MiniMax Code on September 27, 2026, and entered public beta on September 28, 2026. It remains presented as a preview rather than a final general-availability model.

Is MiniMax M3.1 free to use?

Not as a generally free, standalone model. Access is associated with MiniMax Code and subscription plans. A documented unlimited-use promotion ran from October 1 through October 7, 2026, but that offer has expired.

Does MiniMax M3.1 have an API?

Yes. The documentation provides OpenAI-compatible and Anthropic-compatible endpoints using MiniMax-M3.1-Flash-Preview and a MiniMax Subscription Key. Current access should not be assumed to be ordinary open pay-as-you-go API billing.

Is MiniMax M3.1 better than Opus 5.5 or MiniMax M3?

There is no defensible universal answer yet. One independent coding test placed M3.1 below Opus 5.5, while selected creative demonstrations looked competitive. MiniMax M3 has a clearer stable release and a published per-token price, but M3.1 adds newer coding-focused features and reasoning controls. Test both models on your own tasks before choosing.