MiniMax-H3-Image vs HY-Image 3.5: Two Asterisks, Two Very Different Risks

Both of these models come with a qualifier, and the qualifiers are not the same kind of thing. MiniMax-H3-Image, whose weights went up on Hugging Face on 10 November 2026 at 14:50 UTC, is a finished release whose caveat is legal: the MiniMax-H3-Image Community License Agreement excludes the European Union, the United Kingdom, the Republic of Korea and the United States, so a large share of the people who can download it cannot ship it. HY-Image 3.5 is Tencent's current image model and carries no licence obstacle worth mentioning — because there is nothing to license. It is a hosted endpoint, in preview, under the identifier hy-image-v3.5-preview, with no stable non-preview name and no published date for one. One model's risk is a jurisdiction. The other's is a version number that can change under you.

That distinction turns out to decide more of this comparison than the leaderboard does, which is why it is the first thing here rather than a footnote at the end.

What the boards say, and how much of it to believe

Artificial Analysis has both models rated, on both boards, which is more than most matchups in this price band can claim. Read on 14 November 2026:

• Editing — MiniMax-H3-Image 1,138 Elo over 3,137 votes, thirteenth overall; HY-Image 3.5, listed as HunyuanImage 3.5 (Preview), 1,092 over 6,969 votes, twenty-sixth. A gap of forty-six points, with intervals of 1,122–1,154 and 1,079–1,105 that do not overlap.

• Text to image — MiniMax-H3-Image 1,066 over 3,206 votes, twenty-fifth; HY-Image 3.5 at 973 over 6,205 votes, seventy-third. A gap of ninety-three points, again with non-overlapping intervals.

• Price as the boards track it — MiniMax-H3-Image $26.54 per thousand images on the editing board and $21.79 on generation; HY-Image 3.5 $24.00 on both.

• Licence — MiniMax-H3-Image: downloadable, community licence, four excluded jurisdictions. HY-Image 3.5: no weights, no self-hosting, no offline route at any price.

The vote counts are the reason to read those rows carefully rather than suspiciously. HY-Image 3.5 has been on the board since September and has more than twice the samples; MiniMax-H3-Image has been on it for four days. A forty-six-point gap that survives six thousand votes is a stronger claim than one that survives three thousand, and the direction of travel for a new entrant is usually down as the pool widens. So: the ordering is real today and the margin is not yet settled.

The gap inverts between the two boards, and that is the interesting part

MiniMax-H3-Image leads on both, but not by the same amount, and the asymmetry is not symmetric in the way the raw numbers suggest. On editing, forty-six points is a solid but ordinary advantage — a bit over one and a half interval widths. On text-to-image, ninety-three points is a different order of thing: it is the difference between twenty-fifth and seventy-third on a board of a hundred and fourteen models.

Which is the opposite of what the specification sheets would lead you to predict. HY-Image 3.5 is the one that ships with a consistency story — Tencent's claim is that a subject survives repeated edits and re-renders without drifting, which is an editing claim, and its product material leans on reference images rather than prompt fidelity. MiniMax-H3-Image is the one with sixteen megapixels and sixteen positional references, which sounds like an editing configuration too. And yet the boards put Tencent's model relatively stronger as an editor and much weaker as a plain generator, and put MiniMax's model strong on both with a wide lead on the board it was not built for.

The reconciliation is resolution, not capability. Both boards ask voters to prefer one image over another at whatever size the model produced. A model that returns sixteen megapixels when its competitor returns four will win a preference vote on detail more often than it wins on fidelity to the prompt, and the effect grows with the amount of detail in the scene. Whatever the boards are measuring here, part of it is sharpness.

Cost per delivered megapixel, which is the number that actually bills

Per-image pricing hides a factor of four. HY-Image 3.5's ceiling is "up to 2K" in Tencent's own words, which at a square aspect ratio is about 4.2 megapixels. MiniMax-H3-Image returns roughly 16.8 megapixels at every aspect ratio — 4096×4096 at 1:1, 5472×3072 at 16:9, 6272×2688 at 21:9 — and its rate card prices the delivery, not the call: $0.02179 at the 1K tier, $0.02333 at 2K, $0.06879 at 4K list.

The 4K tier is where a sixteen-megapixel image lands, so the comparison is $0.06879 against Tencent's $0.024 for a 2K image internationally. Per image that is nearly three times the price. Per megapixel it reverses: $0.0041 for MiniMax against $0.0057 for HY-Image 3.5, so Tencent's model is 39% more expensive for the pixels you receive. And through 17 November 2026 at 16:00 UTC, MiniMax's 4K tier is discounted by half to $0.03439 — which is $0.00205 per megapixel, 2.8 times cheaper than HY-Image 3.5 for the same delivered area.

That discount has a condition worth reading twice. It applies only when the delivered image is at least 16,600,000 pixels, and to the whole request. A 4K request that comes back at 16.5 megapixels gets no discount at all; the threshold is set just below the model's normal output, which means in practice it is a discount on the size the model is built to produce, and not a lever you can pull on smaller ones.

For an editor the arithmetic is worse for Tencent, because references are billed as input. A 4K edit with one reference image is $0.07312 list, $0.03656 discounted; at 1K it is $0.02654. Three references at 2K is $0.04177. Tencent bills per delivered output image with no separate reference fee that we could find published, so at low reference counts the two rate cards converge and at high counts they diverge — in MiniMax's favour per megapixel, against it per call.

What the preview label costs you in engineering

HY-Image 3.5 has been in preview since September 2026 and Tencent has published no date for the stable release. That is not a scandal; it is how Tencent ships this line, preview first and official version later. But it has a concrete cost that does not appear on any rate card: the identifier is the thing most likely to be superseded, and an identifier embedded in a codebase is a deployment.

The mitigation is unglamorous and it is the whole point. Make the model name a configuration value rather than a constant. Write the reference-image count as a parameter validated against the documented ceiling of five, not as a hardcoded list of five paths. Put the response parser behind one function so that a schema change on Tencent's side touches one file. None of that is specific to this model — it is the discipline that makes any preview-to-official transition cheap, and the teams that skip it are the ones that discover a migration on a Friday.

MiniMax-H3-Image has a mirror-image engineering cost. Nothing about the endpoint will change under you in a way a version pin cannot handle; what will change is whether your organisation is allowed to call it at all. The licence reserves hosted-service rights regardless of revenue, and requires separate written authorisation above a million dollars of annualised revenue. If you are in one of the four excluded jurisdictions, the technical work of adopting it is wasted until a lawyer signs something, and the specific thing to ask for is a written authorisation naming your entity — not an interpretation of the exclusion clause.

Who should pick which

If you need a print-resolution asset out of a single call, MiniMax-H3-Image is the only one of the two that will do it, and at the current discount it is also cheaper per megapixel. That is the case the model was built for and the case where the boards and the rate card agree.

If you need an editor that takes a handful of reference images and holds a subject steady across a run — product variants, a character across a storyboard, a set of listing photos — HY-Image 3.5's documented ceiling of five references and its consistency claim are the relevant specification, and its editing row at 1,092 puts it in the top third of a board where the top ten are almost entirely closed, expensive models. It is also the simpler procurement: a published international rate, no licence review, no jurisdiction question.

If you are in the EU, the UK, Korea or the US and you need both properties, the answer is neither of the two alone. Run HY-Image 3.5 for the editing pass and reach for a downloadable model you are licensed to use when you need the pixels — which, for most teams in those jurisdictions, today means Qwen-Image-2.1 at 1,084 on editing, or FLUX 3 Image Klein at 1,069, both of which are behind MiniMax-H3-Image but ahead of HY-Image 3.5 and both of which you can actually deploy.

Where a router fits

OrcaRouter serves neither model. MiniMax-H3-Image has no route on our catalogue, and HY-Image 3.5 is Tencent's endpoint through Tencent Cloud's TokenHub surface — not ours. What we do carry is the video model from MiniMax's H3 family, minimax/minimax-h3, at the provider's list price with 0% markup passed through.

The reason to care about routing here is that this comparison is a pricing comparison, and pricing comparisons rot. MiniMax's 4K discount expires on 17 November; Tencent's international rate has already moved once since preview. A pass-through arrangement means a vendor price cut is live on our side the same day rather than at the next billing cycle, and a single key across two hundred-odd models means the cost of trying the other one is a request, not a contract. That is not a reason to choose a router over a model. It is a reason not to hardcode a rate card into your own margin calculation.

The two asterisks are not going away this month. MiniMax's four-country exclusion is written into a licence file you can read today, and Tencent's preview designation is the identity the model currently has. Pick the one whose asterisk your organisation can live with, and re-read both after the 17th, when one of the two numbers above stops being true.