MiniMax-H3-Image vs FLUX 3: Coordinates or Pixels — Two Different Ideas of Control

Put these two models in front of the same brief and they will fail in opposite directions. MiniMax-H3-Image gives you sixteen megapixels, sixteen reference images and a six-step sampler, and no way to say where anything goes except in prose. FLUX 3, in its image tier, gives you a bounding box for every element on a 0–1000 grid written as [y_min, x_min, y_max, x_max], lets you edit one box at a time without touching the rest of the frame, and starts its resolution ladder at 768 pixels. One model sells capacity; the other sells direction. If your problem is a product shot that must survive being cropped to a billboard, MiniMax-H3-Image is the only one of the two that can produce the file. If your problem is that the client's logo has to be top-left at exactly 12% of the frame width, FLUX 3 is the only one that will take the instruction.

Both are recent, and neither is a launch in this article. FLUX 3 was announced by Black Forest Labs on 23 July 2026 as a unified video, image, audio and action model; the image tier that this comparison is about went live with a published rate card on 1 October 2026. MiniMax-H3-Image's weights landed on Hugging Face on 10 November 2026 at 14:50 UTC, with a one-line card in M​iniMax's API release notes and no blog post anywhere on the company's site. So this is a comparison between a model that has had six weeks of paying traffic and one that had its first full day in public yesterday. That asymmetry runs through everything below, and it is the reason the honest answer at the end is conditional rather than a winner.

What each one lets you specify

FLUX 3's control surface is spatial, and it is the most opinionated thing about the model. The canvas is always a 0–1000 grid on both axes, normalised, so a layout prompt is resolution-independent: you describe the global scene in a caption, then list elements with an id, a box and a description, and cite the ids inside the caption. The model plans the arrangement — BFL documents an agent mode where an LLM produces the layout — and then every box stays editable afterwards. You can re-describe a box, replace it, or move it, and the regions you did not touch are held fixed. References are numbered tokens, ref_image_0 upward, up to ten of them, and they are composed into one scene rather than stacked as candidates.

MiniMax-H3-Image's control surface is positional and textual. Prompt text goes in as up to 32,000 characters, of which roughly the first 4,096 tokens take effect. References go in as up to sixteen images addressed by upload order — <Picture 1>, <Picture 2> — with no typing at all: nothing in the API tells the model which reference is the character, which is the product and which is the background. The whole edit is expressed in language. There is no second pass, no masked region, no coordinate system.

That difference is not a feature gap in one direction. Sixteen untyped slots is more headroom than ten numbered references, and a caption that describes a relationship — “the woman from picture 3, wearing the jacket from picture 7, in the room from picture 1” — carries information a bounding box cannot. But a bounding box carries precision a caption cannot, and MiniMax-H3-Image has no equivalent.

The numbers, dimension by dimension

• Output ceiling — FLUX 3 Image: 4K, published as 5456×3072, about 16.8 MP. MiniMax-H3-Image: about 16 MP at every aspect ratio, 4096×4096 at 1:1 up to 6272×2688 at 21:9, capped at 20 MP by the API.

• Output floor — FLUX 3 Image: 768×768. MiniMax-H3-Image: the card names roughly 4 MP as the practical bottom, with no small-image tier at all.

• Price at 1K — FLUX 3 Image: $0.048. MiniMax-H3-Image: $0.02179.

• Price at 2K — FLUX 3 Image: $0.10. MiniMax-H3-Image: $0.02333.

• Price at 4K — FLUX 3 Image: $0.607. MiniMax-H3-Image: $0.06879 list, $0.03439 under a discount that expires 17 November 2026 at 16:00 UTC.

• References — FLUX 3 Image: up to 10, numbered, free of charge. MiniMax-H3-Image: up to 16, positional, billed as input image tokens at $1.50 per million.

• Editing — FLUX 3 Image: box-scoped, multi-box per request, untouched regions preserved. MiniMax-H3-Image: whole-image, prompt-described, up to 16 references in one call.

• Weights — FLUX 3 Image: not released; commercial weights licensing is sold to companies at scale through BFL sales, and the FLUX 3 Dev backbone remains “coming soon.” MiniMax-H3-Image: downloadable today, 68.65 GB plus a 66.7 GB encoder borrowed from MiniMax-H3.

• Sampling — MiniMax-H3-Image: fixed at six steps, no classifier-free guidance, other values rejected with HTTP 400. FLUX 3 Image: no published step count or guidance control; the image tier is API-only.

• Independent score — FLUX 3 Image: Elo 1116 on Artificial Analysis's text-to-image board, 16th of 108, over 6,457 samples; Elo 1119 on the editing board, 17th, over 5,256 samples. MiniMax-H3-Image: no entry on either board, and no vendor evaluation table either.

The price gap is bigger than the resolution gap, and it points the other way

At 1K, MiniMax-H3-Image costs $0.02179 and FLUX 3 Image costs $0.048 — a little over half. At 2K the ratio widens to roughly four to one. At 4K it is nearly nine to one: $0.607 against $0.06879, or twelve to one against the discounted $0.03439. That curve is the single most consequential fact in this comparison, and it is worth understanding why it is shaped that way rather than assuming one vendor is gouging.

FLUX 3 Image's ladder — $0.041, $0.048, $0.07, $0.10, $0.607 — is close to proportional to pixels up to 2K and then jumps hard at 4K. M​iniMax's ladder is nearly flat: $0.02179, $0.02333, $0.06879. Sixteen times the pixels for a little over three times the money. M​iniMax meters tokens and the token count does not scale with area the way a naive model would predict; BFL appears to price the 4K tier as a premium product. Whatever the reasoning on either side, the practical consequence is that at the resolution both models were built for, the open-weight one is nearly an order of magnitude cheaper.

That is also where the comparison stops being symmetrical. If you only ever generate 1K images, the price difference is two and a half cents and FLUX 3's layout control is probably worth more than the saving. The 4K tier is the only tier where the arithmetic changes a decision, and it is exactly the tier where M​iniMax's discount is temporary.

What six weeks of traffic buys, and what a day does not

FLUX 3 Image has been scored. Artificial Analysis places it 16th on text-to-image at Elo 1116 over 6,457 samples, and 17th on editing at Elo 1119 over 5,256 — mid-table, above FLUX.2 [max] and below most of the current Nano Banana and MAI field, at a listed $48 per thousand images. You can put that next to a competitor's number and make a decision. You can also watch it move as more samples arrive, which is what happened to several models above it.

MiniMax-H3-Image has nothing comparable. Not a low score — no score. The vendor's card claims “the strongest instruction following among open-source models” without a table, and the showcase grid on the model page is the only evidence of output that exists. One day in public is not enough time for a blind arena to accumulate samples on a model that takes ten seconds a generation, so the absence is expected rather than suspicious. It is still an absence, and it means the argument for MiniMax-H3-Image on quality is currently an argument from architecture and from price.

The architecture argument is not empty. A 33.1B dense diffusion transformer distilled to six unguided steps is a deliberate choice about serving cost, and six steps is why a 16 MP generation is a request rather than a batch job. The card's own framing — 16 MP “at a practical computational cost” — is a FLOPs claim, and it is consistent with everything in the repository. It just is not a quality claim, and the two get conflated in every launch post.

Where each one breaks in production

FLUX 3's limit is its smallest output. A 768-pixel floor and a 4K ceiling mean the model is comfortable in the range where most commercial work happens, and it will not give you a 6000-pixel-wide plate. It is also closed: you cannot download it, you cannot pin a checkpoint, and BFL's licensing page sells self-hosting to companies at scale rather than to individuals. If BFL changes a rate or retires a tier, your pipeline changes with it. Six weeks of traffic is a track record, not a guarantee.

MiniMax-H3-Image's limit is the licence, and it is stricter than “open weight” suggests. The MiniMax-H3-Image Community License Agreement excludes the European Union, the United Kingdom, the Republic of Korea and the United States by definition; requires separate written authorisation above one million US dollars of annualised revenue; requires the same for any hosted service with no revenue floor; requires the model's name displayed prominently on commercial interfaces; and forbids using outputs to train other models. If you are in one of those territories and building a product, the open weights are not available to you on these terms, and the hosted API — which M​iniMax sells — is the path that remains.

There is a third limit that both share and that no launch post mentions: the evaluation loop. Neither model gives you a way to know whether a given prompt will land except by generating it. FLUX 3's boxes reduce the space of possible failures to placement, which is checkable in code. MiniMax-H3-Image's failures are semantic, which means a human has to look. For a pipeline generating thousands of images a day, that difference in verification cost can exceed the difference in generation cost.

Which one to call

If the deliverable is large — a print plate, a background that will be cropped to three aspect ratios, a 4K hero that has to hold detail at 100% — MiniMax-H3-Image is the only side of this that produces the file, and at $0.06879 a frame it is also the cheaper one by a wide margin. Take it, budget at list price rather than the discounted number, and read the licence before you build a product on it.

If the deliverable is compositional — a layout that has to be exact, an edit that has to leave everything else alone, a brand system where placement is the whole point — FLUX 3 is the only side with the controls, and it is scored, which means you can predict what you are buying. Two and a half cents more at 1K is not the deciding number there.

If you need both, the pattern that actually works is to split the job rather than pick a side: compose and edit at the resolution FLUX 3 handles, then upscale or regenerate the approved frame at 16 MP. That is two calls and two bills instead of one, and it is still cheaper than discovering after a hundred generations that your model cannot hold a logo in the corner.

Neither model is on OrcaRouter's catalogue today — FLUX 3 Image is a direct Black Forest Labs call, and MiniMax-H3-Image is M​iniMax's own API or your own GPUs — so nothing here is a routing recommendation. What we do carry is the video half of M​iniMax's H3 family, minimax/minimax-h3, alongside the image routes we do host, under one key with 0% markup on provider list price. The reason that matters in a comparison like this one is failover: when a 16 MP request takes longer than your client's timeout, a second route under the same interface is worth more than a better per-image price. Long, billed-on-timeout generations are precisely the workload where a fallback earns its keep, and M​iniMax charges for a timeout rather than refunding it.

The question that will settle it

The thing missing from this comparison is not a spec. It is a sample count. When MiniMax-H3-Image appears on an image arena with a few thousand blind votes, the price gap stops being the main argument and becomes a trade: some amount of measured quality for some amount of money. Until then, FLUX 3 is the model you can reason about and MiniMax-H3-Image is the model you have to test, and the cost of testing it is low enough — a few dollars and an afternoon — that anyone with a 4K deliverable should simply run the test rather than wait for someone else's.