MiniMax H3 Content Errors: Why Hailuo Blocks Videos That Already Finished

TL;DR: MiniMax runs two separate content checks, and they fail very differently. Code 1026 rejects your prompt before anything renders. Code 1027 lets the job run all the way through, then blocks the finished video. In the Hailuo app both surface as the same vague line, which is why "just try again" fixes one and does nothing for the other. The check lives in MiniMax's hosted platform, not in the H3 weights, so the same model on another stack behaves differently.

You write a prompt. Hailuo accepts it. The progress bar fills. And at the very end, instead of a video, you get a short apology and no output.

That last-step failure confuses people more than an outright rejection does, because everything up to that point looked fine. The app doesn't explain what happened, and the obvious response, hitting regenerate, often produces the same result.

There are two different failures wearing the same coat.

The exact strings you'll see

In the Hailuo app and on the web, the message is short and gives you almost nothing:

Content generation error, please regenerate

Users also report it landing specifically "at the final step of generation while using images," which is the detail that matters and the one the message omits.

Through MiniMax's API you get something far more precise. Their published error-code reference defines these:

The official remedy printed next to both is the same sentence: "Please change your input content." For 1027 that advice is close to useless, since the thing that got flagged wasn't your input.

1026 returns HTTP 422. It arrives fast, because nothing rendered.

Why one of them wastes your time and the other doesn't

This is the part worth internalising.

1026 is cheap. The filter reads your prompt, decides it doesn't like a word or a concept, and stops. You lose a few seconds. Rewrite, resubmit, move on.

1027 is expensive. Your prompt passed. The model generated a complete video. Then a second check looked at the frames that came out and rejected them. You waited through the whole render for nothing.

That asymmetry explains the most common complaint about Hailuo: that regenerating doesn't help. For a 1026 it usually does, because a reworded prompt is a different input. For a 1027 you're feeding the same accepted prompt back into a model that will produce something very similar, which the same output filter will then judge the same way. Same input, same outcome, more waiting.

One user description of the newer platform captures the mechanic exactly: the prompt is accepted, and after the video finishes it goes into moderation to be checked against policy.

Why it fires on prompts with nothing wrong in them

Output filters don't read your intent. They score pixels.

A model that scores finished frames has to make a judgement call on ambiguous imagery, and the safe direction for the operator is always to block. Reports from Hailuo users include beach and pool scenes getting flagged with no swimwear mentioned anywhere in the prompt. Nothing in that request is objectionable. The filter simply saw skin and water and rounded up.

You can't prompt your way around this reliably, because the trigger isn't in the prompt. It's in whatever the model happened to render.

Where the check actually lives

Not in H3 itself.

MiniMax open-weighted part of H3, and those weights don't carry a content filter. The 1026 and 1027 checks belong to MiniMax's hosted platform, the layer wrapping the model on their API and in the Hailuo app. Take the same model, serve it somewhere else, and that layer isn't in the path.

We tested this directly rather than assuming it. Running a set of deliberately boundary-adjacent prompts through H3 on an open-weights serving channel, all five completed and none were refused. Same model, different serving stack, different behaviour. The full method and results are in our H3 moderation test.

This is the same shape as Wan 3.0's DataInspectionFailed, where Alibaba's compliance layer screens input and output separately, and GPT Image 2's two-stage filter. Three different vendors, one pattern: the restriction is a property of the platform you rent, not the model you think you're using.

Running H3 yourself, and what it costs

Since the weights are public, you can sidestep the hosted filters entirely by running H3 on your own machine through ComfyUI. No API, no platform layer, no 1026 or 1027.

The hardware is the catch, and it's a real one. MiniMax publishes no official minimum, but the practical picture from people running it looks like this:

The weights themselves run about 42.5 GB on disk for the smallest usable set, down from 123.6 GB at full precision.

For anyone who already owns a 24 GB card, this is a good answer. For everyone else it means buying a dedicated GPU to avoid an error message, and the maths rarely works out. A single upfront hardware purchase to fix an occasional blocked render is the expensive way to solve a cheap problem.

What you can do

If you're getting 1026, rewrite. Strip the specific noun or scenario the filter is likely reacting to and resubmit. This one responds to editing.

If you're getting 1027, rewriting the prompt is mostly guesswork, because you're trying to steer pixels you haven't seen yet. You have three real options: change the scene enough that the render is unlikely to land in ambiguous territory, run the weights locally if you own the hardware, or generate on a stack that doesn't sit behind MiniMax's platform filter.

What that looks like on VidCella

We serve MiniMax H3 through its open-weights channel, so MiniMax's hosted moderation layer isn't in the request path and the 1026 and 1027 codes don't apply here.

That isn't the same as saying anything goes. We run our own safety checks, and they're built to be precise, so ordinary creative work isn't caught by a filter reacting to skin tones and water. What you don't get is a generation that completes and then disappears because a second model disagreed with the first.

The practical specs: 480p or 768p output, 5 to 15 seconds per generation, with native audio. Text-to-video and image-to-video run 12 credits per second at 480p and 24 at 768p. Reference-to-video is 15 and 30. Pay per generation, no subscription.

FAQs

What does error 1026 mean on MiniMax? Your prompt was rejected by the input content check before any video was generated. MiniMax labels it input new_sensitive and returns HTTP 422. Rewording the prompt usually clears it.

What does error 1027 mean? The video generated successfully and was then rejected by the output content check. MiniMax labels it output new_sensitive. Because the block is on the rendered frames rather than your text, resubmitting the same prompt tends to reproduce it.

Why does Hailuo fail at the last step instead of rejecting my prompt? Because the output check runs after rendering. A prompt that passes the input filter can still produce frames the output filter rejects, and that decision can only be made once the video exists.

Does regenerating fix it? For an input-side block, often. For an output-side block, rarely, since the same accepted prompt will produce similar frames for the same filter to reject again.

Is MiniMax H3 itself censored? The open weights don't carry the filter. The 1026 and 1027 checks come from MiniMax's hosted platform, which is why the same model can behave differently depending on who is serving it.

Can I avoid this by running H3 locally? Yes, the hosted checks aren't in the path when you run the weights yourself. It takes roughly 42.5 GB of disk and a 24 GB GPU for a comfortable setup, so it's only worth it if you already have the hardware.

Related reading

• Is MiniMax H3 uncensored? We tested it — our own boundary test on the open-weights channel
• Wan 3.0 "Inappropriate Content" error — Alibaba's version of the same two-stage check
• Why GPT Image 2 blocks your prompts — the same pattern on OpenAI's stack
• Open weights doesn't mean uncensored — where content filtering actually sits