Notes on the limits of generative video APIs that reject rather than degrade — the values that come back as an error instead of a slightly worse clip.
Soft limits are self-correcting. Ask for something the model is bad at and you get a bad clip; you look at it, you change the prompt. The feedback is immediate and the cost is one generation.
Hard limits do not work like that. A four-second request on a model whose floor is five does not return a short clip — it returns nothing. A pair of legal-sized images that together exceed a shared request budget does not get compressed — it gets refused. These fail at the boundary, often on a path nobody tested, and they look like the integration is broken rather than misconfigured.
They are also the ones missing from most write-ups, because they only surface once you have built something.
The team behind minimaxh3max.video, an independent third-party interface to MiniMax H3 Max. We are not affiliated with MiniMax.
That is a commercial interest, and it is the reason every number here carries a source and a change date rather than an assertion. You should be able to check any figure rather than take our word for it.
The MiniMax H3 Max specifications page lists eighteen limits, each read from the constant the generator actually enforces rather than copied off a model card, and each with the date it last changed.
That construction is deliberate. It means the page and the Generate button read the same value, so the page cannot drift from the product — if a number there is wrong, the product is wrong in the same way. Model specifications move, and a rate card or limit table with no date on it is a claim about a moment that has already passed.
If you find an error on any page here, it is worth telling us.