MiniMax-H3-Image vs Nano Banana 2: Free Weights, Closed Borders

The surprising thing about MiniMax-H3-Image is not its score. It is the map in its licence. The checkpoint went up on Hugging Face on 10 November 2026 — a 68.65 GB repository, 66.28 GB of it the diffusion transformer, plus a 66.7 GB text encoder it borrows from the H3 video model — and the file that governs it, the MiniMax-H3-Image Community License Agreement, grants rights "worldwide, excluding the Excluded Territories," which it then defines as the European Union, the United Kingdom, the Republic of Korea and the United States of America. Nano Banana 2, G​oogle's gemini-3.1-flash-image, has been generally available since 28 May 2026 and asks nothing of you but a credit card. So the honest framing of this matchup is not open versus closed. It is a model you can download but may not legally deploy where you live, against a model you can never download but can call from anywhere. Everything else — the sixteen megapixels, the price ladder, the leaderboard positions — is secondary to that, and the rest of this piece is what changes once you accept it.

What the licence actually permits

Read on 14 November 2026, Section IV of the M​iniMax licence is short and unusually specific. Two things trigger a requirement to obtain prior written authorisation through M​iniMax's licence console: annualised revenue above US$1 million for you or your affiliates, or providing a hosted service to third parties. Neither trigger charges a royalty. Both require you to ask. Section IV also imposes an attribution duty — any commercial product using the model must "prominently display 'MiniMax-H3-Image'" in its interface, which is a different obligation from the usual "built with" badge because it names the model, not the company.

The territory clause is the harder one, and it is not a nuance. The grant itself is scoped to the applicable territory, and the agreement says plainly that use outside it "is not authorized by this Agreement." That is not a restriction on redistribution; it is a restriction on running the weights at all in the four listed jurisdictions. M​iniMax's own card points anyone in those regions at a separate licensing conversation, mentioning "robust controls and guardrails" as part of what such an arrangement would involve. For a studio in Berlin, a startup in Seoul, or — more to the point for most readers of an English-language model blog — anywhere in the United States, the practical answer is that open weights are a research artefact you can inspect, benchmark and write about, and not a production dependency.

Nano Banana 2 has no equivalent puzzle. It is a G​oogle Cloud product: you call it, you get an image, you ship it, and the terms are the standard API terms. There is no revenue threshold, no territorial carve-out, no attribution clause, and no hosted-service negotiation. What you give up for that simplicity is the file. If G​oogle changes the price, retires the identifier, or has an outage, you have a migration ticket rather than a fallback.

That symmetry — one model gives you the artefact with an unusable licence, the other gives you a usable licence with no artefact — is the actual decision. Most comparison pages skip it because the licence file is boring and the benchmark table is not.

Where the two models cross on the boards

Artificial Analysis now scores both of them on both of its image arenas, which was not true when MiniMax-H3-Image shipped. Read on 14 November 2026, the numbers describe a genuine crossing rather than a gap:

• Text to image — Nano Banana 2 sits 11th of 114 rows at Elo 1,143 (95% interval 1,133–1,153). MiniMax-H3-Image sits 25th at 1,066 (1,051–1,080). That is 77 Elo, and the intervals do not come close to touching.

• Image editing — MiniMax-H3-Image sits 13th of 84 at Elo 1,138 (1,122–1,154). Nano Banana 2 sits 17th at 1,126 (1,114–1,138). M​iniMax is ahead by 12 Elo; the intervals overlap by a few points, so treat this as a narrow lead rather than a rout, but it is the direction the board shows.

• The Highspeed variant — MiniMax-H3-Image Highspeed, which exists only on M​iniMax's own API, is ranked separately: 24th on text to image at 1,072 and 14th on editing at 1,134. It costs roughly 1.7× the standard tier per generation at 4K and does not buy a better score on either board. It buys latency, which the boards do not measure.

Artificial Analysis's own summary of the results, posted on 13 November 2026, adds the framing M​iniMax would want: on its boards, MiniMax-H3-Image is the highest-ranked row carrying the open-weights marker in both categories. That claim is theirs and it depends on their classification of which rows count as open weights — several of the models above H3-Image on text to image, including the FLUX 3 Image row at 1,119, are not obviously downloadable, so the marker is doing real work. Quote it as a vendor-of-the-leaderboard statement rather than a fact you can verify from the weights alone.

The shape to hold onto: M​iniMax's model is better at the edit task and clearly worse at the blank-page task. That matches its design. Sixteen reference slots and a sampler tuned for reference-heavy work push the editing board; a six-step distilled sampler with no guidance is a control story rather than a fidelity story, and the text-to-image gap is what that costs.

The price ladder, and a discount with three days left

M​iniMax bills by token rather than by image, which makes the headline number misleading until you read the worked examples. Standard tier: $2.30 per million input text tokens, $1.50 per million input image tokens, $3.80 per million output image tokens. Highspeed: the same text rate, $3.00 on image input, $7.60 on output.

• Text to image, standard — $0.02179 at 1K, $0.02333 at 2K, and $0.03439 at 4K while the launch discount holds, against a list price of $0.06879.

• Edit with one reference — $0.02654 at 1K, $0.02658 at 2K, $0.03656 at 4K (list $0.07312).

• Edit with three references — $0.03563 at 1K, $0.04177 at 2K, $0.03957 at 4K (list $0.07914).

• Highspeed, text to image — $0.03529 at 1K, $0.03622 at 2K, $0.06501 at 4K (list $0.13001).

The discount that halves the 4K column runs until 17 November 2026, 16:00 UTC, and applies only to delivered images of at least 16,600,000 pixels. After that the 4K rate roughly doubles while every lower tier stays put — which changes the shape of the recommendation, because the flat ladder is the reason to use this model at all. At list price, a 4K M​iniMax image costs about the same as a 4K Nano Banana 2 image and no longer undercuts it.

G​oogle's side is a clean per-image table: 1,120 tokens for a 1K image and 2,520 for 4K, at $60 per million image output tokens. That resolves to $0.045 at 512px, $0.067 at 1024×1024, $0.101 at 2048×2048 and $0.151 at 4096×4096, with text and thinking billed separately at $3 per million output tokens and inputs at $0.50 per million. There is no batch tier listed for this model, unlike its Lite sibling. Note the asymmetry in what "at resolution" means: a Nano Banana 2 image at 4K is 4096×4096, and a M​iniMax image at 4K is also 4096×4096, so the two ladders compare directly at the top and diverge at the bottom, where G​oogle sells a 512px tier that M​iniMax does not.

Ten thousand images a month, one reference each, 2K output: about $266 on M​iniMax standard at the current rate and about $1,010 on Nano Banana 2. That is the one place where the numbers are not close, and it is also the place where the licence is most likely to stop you.

What one can do that the other cannot

The capability lists are less symmetrical than the scores. Both models generate and edit through a single endpoint pair, both accept reference images and both put out multi-megapixel files. Past that they diverge on nearly every axis:

• References — MiniMax-H3-Image takes 1 to 16 images, addressed positionally in the prompt as <Picture 1> through <Picture 16> in upload order. Nano Banana 2 takes 14, but splits them: up to 10 object references for fidelity and up to 4 character references for consistency.

• Output — M​iniMax claims roughly 16 megapixels at every aspect ratio, from 4096×4096 square to 6272×2688 ultrawide, with a 20-megapixel API ceiling and no upscaler in the path. Nano Banana 2 tops out at 4096×4096 but adds the extreme 1:4, 4:1, 1:8 and 8:1 shapes and a 512px tier for thumbnails.

• Grounding — Nano Banana 2 can ground a generation in G​oogle Search or Image Search results. MiniMax-H3-Image has no retrieval of any kind; everything comes from the prompt and the references.

• Provenance — every Nano Banana 2 output carries a SynthID watermark and C2PA metadata. The M​iniMax repository and licence state no watermarking scheme, which is either a feature or a compliance problem depending on your jurisdiction, and in the EU it is the latter.

• Reasoning controls — Nano Banana 2 exposes thinking at minimal or high, defaults to minimal, and cannot have thinking switched off through the API; those tokens are billed whether you read them or not. M​iniMax exposes no reasoning surface at all, just a fixed sampler.

• Sampler — M​iniMax is distilled to six steps with no classifier-free guidance. The API returns HTTP 400 if you pass any other step count or guidance value, and it rejects negative prompts. You tune the prompt and the references, and nothing else. G​oogle's model takes the usual parameter surface.

The sampler point cuts both ways. It makes M​iniMax's output exactly reproducible on your own hardware, which is a real property for regression testing, and it removes the standard quality dial that a team used to tuning CFG would reach for first when an edit comes out wrong.

If you were going to self-host, read this first

The reason most people want open weights is not ideology; it is the option to stop paying per image. On this model that option is expensive in a way the download size hints at and the announcement does not. The transformer is 66.28 GB in BF16, the VAE 2.34 GB, and the text encoder is a separate 66.7 GB pull — the full Qwen3-VL from MiniMax-H3, unchanged, whose 50th-layer hidden states the DiT reads. That is roughly 135 GB of weights before you have a single image, and the model is a 33.1-billion-parameter dense transformer, not a mixture-of-experts with a cheap active path.

Against that, Qwen-Image 2.1 — the other open option people ask about — downloads at roughly a quarter of the size. If your plan is a single workstation or one rented GPU, the arithmetic on this checkpoint is the first thing to run, not the last. M​iniMax's own card offers no latency figure and no hardware recipe, so there is nothing in the release material to tell you whether six steps at sixteen megapixels fits your card.

Which one, given all of that

If you are outside the EU, UK, South Korea and the United States and you need volume at 4K with a lot of references, MiniMax-H3-Image is the better and cheaper editing model right now, and the 12-Elo editing lead is small enough that you should test it on your own prompts rather than trust the board. If you are inside those territories, the licence has already made the decision for you, and the interesting question becomes whether M​iniMax will license it — which is a commercial conversation, not a technical one.

If you need text rendering, search grounding, provenance marking, or an unbounded caller-side contract, Nano Banana 2 wins on capability and loses on price, and its 77-Elo text-to-image lead is the largest single number in this comparison.

Neither model is on OrcaRouter's catalogue as of 14 November 2026, and this page will not pretend otherwise. What the platform does carry in the image line is the O​penAI and G​oogle image models that are actually callable today, at provider list price with 0% markup passed through — which means that when M​iniMax's 4K discount expires on 17 November, or when G​oogle adjusts a rate, the number on our catalogue moves the same day rather than at the next contract renewal. The other reason the routing layer matters here is failover: a model with no independent latency figure and a six-day-old leaderboard entry is exactly the kind of dependency you want behind automatic failover rather than wired directly into a production path.

What would change this verdict is a licence amendment. M​iniMax has shipped, praised and then relicensed models before, and the territory clause reads like a compliance default rather than a considered restriction. If the Excluded Territories list shortens, the editing side of this comparison becomes a straightforward recommendation on price and score. Until then, the open model is open in the sense that you can read it, and closed in the sense that decides whether you can use it.