MiniMax AI Detector

Here you can check English text with a free MiniMax AI detector. Paste a passage of 100 words or more into the checker above and V4, our current engine, scores it for the writing patterns of MiniMax and other AI models: an AI-writing signal score from 1 to 99, a stability range and a verdict of Likely human-written, Uncertain or Likely AI-written. Each result also shows how often human writing of that length was wrongly flagged in our tests. No sign-up, no cost, and your text is not kept.

MiniMax writing was in V4’s training data, and in our sealed test the frozen engine caught 331 of 333 MiniMax texts of 150 words or more, a larger sample than most families had. The sections below explain how to use the MiniMax checker, what a result can and cannot show, and how the test was run.

Can a MiniMax AI detector confirm that MiniMax wrote a text?

No. It can confirm that a passage reads like AI writing, which raw MiniMax output almost always does, but it cannot point to MiniMax as the author. V4 learned from nine model families together and gives a single score for AI-writing patterns; it never names a model.

• It does score English text of 100 words or more and give one of three verdicts.
• It does catch unedited MiniMax output: 99.40% at 150 words or more in the sealed test.
• It does not say whether MiniMax, Kimi, ChatGPT or another assistant wrote a passage.
• It does not prove authorship. Rewording, cutting and translating all change the score.

What the MiniMax AI checker does

The checker runs one fine-tuned transformer text classifier. It reads your passage in overlapping windows, averages them into one internal value and maps that onto the 1–99 scale. It was trained once on texts from nine model families, MiniMax among them, plus a large body of human writing from 2021 or earlier, and then frozen. Nothing you paste changes it.

It does not look your text up anywhere, check it for plagiarism or judge it line by line. The score is about the passage as a whole. How the detector works describes the engine in more depth.

The score is not the share of the text written by AI, and it is not a probability. A result alone does not prove who wrote a text.

Where MiniMax text comes from

MiniMax is less of a household name than ChatGPT, but its models are available through many apps and developer platforms, and text from them can turn up in the same places as any other assistant’s: essays, product copy, support replies and social posts. People rarely say which assistant they used, and a detector cannot tell either.

That is why the useful question is not “was this MiniMax?” but “how much of this reads like unedited AI writing, and what else do I know about how it was written?” The checker answers the first part; drafts, sources and a conversation with the writer answer the second.

How to check MiniMax text

1. Copy the original. Use the text as it was submitted or published.
2. Paste 100 words or more into the checker; 150 words or more gives a more reliable reading.
3. Run the check and read the verdict, score and stability range together.
4. Read the evidence line for the error rate measured on human texts of that length.
5. Look at the bigger picture before deciding anything: drafts, version history and the writer’s account.

You can compare up to 3 passages side by side and save any result as a PDF.

Reading a MiniMax detector result

• Likely human-written (1–34): the passage looks like most human writing of its length. Edited AI text can land here as well.
• Uncertain (35–64): above typical human writing, below the AI line. Treat it as unresolved.
• Likely AI-written (65–99): at or above the line, with a Moderate signal chip for 65–84 and a Strong signal chip for 85–99.

The AI line is 65 for texts of every length. Each length band has its own frozen internal threshold, and the score is mapped so that 65 sits on it. The stability range shows how far the score moves when the source families behind the score map are resampled; the verdict comes from the frozen line. Read understanding results for the whole card.

Expect raw MiniMax text to score high. When a MiniMax passage lands lower, the usual reasons are editing or a very short sample, not human authorship.

How V4 was trained on MiniMax text, and how well it detects it

MiniMax writing is one of the nine families V4 was trained on. To measure how well the frozen engine handles MiniMax texts it had never seen, we scored a sealed set a single time, on 17 October 2026.

The MiniMax sample was built between 10 and 13 October 2026: 35 task formats such as reports, reviews, letters, explainers and dialogue, each requested in 6 styles of writing, with half of the subjects seeded by real human passages. The texts came through the Together API at the provider’s default settings, as a default-tier model with a larger target count, which is why its sample is bigger than most. Reasoning output was deleted and duplicates removed, leaving one final answer per prompt.

What the test found:

• A large, steady sample. 331 of 333 texts of 150 words or more were caught (99.40%), and 53 of 54 at 100 to 149 words (98.15%).
• Length helped, as usual: 98.15% at 100 to 149 words, 98.82% at 150 to 299, 99.20% at 300 to 599 and 100.00% at 600 or more.
• Without MiniMax in training, in the development test, 98.00% of its texts were still caught.
• The previous engine caught 33.63% of the same texts, about one in three.
• Rewriting tricks were measured on the whole development set, not per family: after synonym substitution, the weakest point, 97.40% of AI texts were still caught.

The evidence page has every family, length band and human source, and the methodology explains how the sets were built.

One real example

Development-set example, kept out of the sealed test: MiniMax produced a high-school essay of 258 words that the engine scored 92, stability range 89–95, zone Likely AI-written. Its opening 61 words read:

Cognitive biases are patterns of thinking that happen in our brain when we process information and make decisions. Everyone have these biases, even if they are very smart or educated. The human brain try to simplify the huge amount of information we receive every day, but sometimes this simplification make us think in ways that are not always correct or logical.

Edited MiniMax text and false positives

Every MiniMax text in the test was untouched model output. Paraphrased, edited or translated versions were not scored and typically come out lower, so the figures above do not describe them, and a low score on such text is not evidence of human writing.

Human writing can be wrongly flagged as well. In the sealed test, human texts from sources the engine never trained on reached 65 in 4 of 2,863 cases at 150 words or more (about 1 in 700) and 11 of 1,136 at 100 to 149 words (about 1 in 100). For texts of 600 words or more, none of 970 did. Learners of English were flagged about 3 times as often as native writers, and formal speeches in 17 of 931 cases. If this happened to your writing, what to do if your text is flagged can help.

Using a MiniMax checker fairly

Use a score to choose what deserves a closer read, never as a verdict on a person. Check the rules that apply, ask the writer how the text was produced and look at drafts and sources. Do not grade, reject or accuse anyone on a score alone, especially for short texts and writers working in a second language.

MiniMax AI detector FAQ

Other AI model guides

Each guide uses the same checker and reports its family’s own results. Browse all model guides, or compare with Kimi, GLM, the Llama AI detector and the Claude AI detector.

Tested and written by the AI Detector Checker team. Last tested: 17 October 2026.