MiniMax-H3-Image vs MiniMax-H3: One Is a Video Model, the Other Is Its Still-Frame Cut

The most honest thing about MiniMax-H3-Image is its file listing. Sixty-six gigabytes of diffusion transformer, 2.34 GB of VAE, twelve megabytes of per-step sampling heads — and no text encoder anywhere in the repository. The encoder is not missing. It is borrowed: model_index.json reaches into MiniMax-H3, the 33B omni-modal video model M​iniMax open-sourced on 31 July 2026, and loads text_encoder/ from there, 66.7 GB of Qwen3-VL, unchanged. If MiniMax-H3 is already on your disk, the image model costs nothing extra to encode a prompt with. If it is not, your first still frame costs 135 GB of downloads, and the second model in this comparison is doing real work for the first one.

That one line of configuration is the relationship between the two models, and it explains most of the practical differences between them. MiniMax-H3-Image is MiniMax-H3 with the audio branch removed, the output head retargeted from a latent video to a latent image, and the whole thing distilled down to six sampling steps. Nothing about the text side changed. The DiT even reads the same 50th-layer hidden states off the same encoder, which is why the tokenizer and processor folders are byte-identical and why the image model inherits H3's English-language prompt handling without a line of new training on that path.

What makes this worth writing about now is the manner of the release rather than its contents. M​iniMax has not published a blog post for the image model. The company's blog and news index carry nothing about the image model — the newest item there is the M3.1 Flash post of 19 October — and the model pages still lead with H3 as a video system. The only vendor-side acknowledgement we could find is a card dated 10 November 2026 in the API documentation's model release list, reading “A leading open-weight image model. From ideas to new worlds” — a sentence with no specification, no benchmark and no date of its own. The Hugging Face repository was created on 3 November, and the weights landed on 10 November at 14:50 UTC. Two README revisions and a vLLM-Omni deployment guide followed within the hour. That is a quiet ship: the artefacts are real and downloadable, the announcement is not. Everything below is read off the repository, the licence file and M​iniMax's own documentation, and where a number is the vendor's own unreproduced claim we say so.

The two models, in the terms each one actually sells on

MiniMax-H3 is a general-purpose omni-modal generator. It takes text, images, video and audio as one packed context and emits 4 to 15 seconds of video at 24 frames per second with native 32 kHz stereo audio, at 768p natively, with a separate regeneration pass — H3-Regenerate-2K — that feeds the result and the original context back through the model to reach 2K. Two input variants matter: H3-Base-FL2VA takes zero, one or two images as first/last frames, and H3-Base-Ref2VA takes up to 9 images, 3 video clips and 3 audio clips, capped at 12 files overall. A hosted preprocessor called H3-Context-IR parses the mixed input and rewrites it into a structured form, and M​iniMax is explicit that it is not part of the open release because it depends on several hosted models.

MiniMax-H3-Image takes text and up to 16 reference images and emits one image. The diffusion transformer is 33.1B parameters — a dense single-stream design, 50 layers, hidden size 5,376, 56 attention heads, FFN width 14,336, 3D multimodal RoPE, BF16 weights with FP32 embedding, timestep and output layers. The image VAE is a CNN encoder with a 24-layer ViT decoder, 16× spatial downsampling and 24 latent channels; the DiT patchifies that latent 2×2, so one transformer token covers a 32×32-pixel block. Output area runs to 20 megapixels through the API, and the card's headline claim is roughly 16 megapixels at every aspect ratio — 4096×4096 at 1:1 up to 6272×2688 at 21:9, without an upscaler.

Read those two paragraphs side by side and the fork is obvious. The video model spends its capacity on time and on a second modality. The image model spends it on space. Same trunk, different objective.

Six steps and no classifier-free guidance is the actual engineering

The part of MiniMax-H3-Image that is not just subtraction is the distillation, and it is enforced rather than recommended. The checkpoint runs six sampling steps — six NFEs — with no classifier-free guidance at all. The served API does not treat those as defaults: num_inference_steps must be omitted or 6, guidance_scale must be omitted or 1, and any other value returns HTTP 400. Negative prompts are rejected outright. A separate file, pdd_fused_heads.safetensors, carries per-step output heads that replace the transformer's output layer, with the schedule in pdd_config.json; that is 12 MB of the 68.65 GB total doing a disproportionate amount of the work.

This is the difference between a model you can put behind an interactive surface and one you cannot. A guided 30-step diffusion model at 16 megapixels is a batch job. Six unguided steps at the same resolution is a request. M​iniMax's own framing — “16-megapixel images at a practical computational cost” — is a claim about FLOPs per pixel, and the six-step schedule is how the claim is cashed. Whether it holds at the quality level the showcase images suggest is a question no third party has answered; there is no Artificial Analysis entry for MiniMax-H3-Image on either image board as of 10 November 2026, and no vendor evaluation table in the card.

For local serving the framework story is unusually complete for a model this young. SGLang is the reference path with a documented cookbook; vLLM-Omni publishes a recipe with per-hardware configurations and measured latency, naming NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X; diffusers and ComfyUI both ship support. Sequence parallelism is a flag — --ulysses-degree on SGLang, --usp on vLLM-Omni — and a single GPU is enough to start. The text encoder is the real memory cost, and it is the video model's text encoder, which is the point of this whole comparison.

Two meters, and neither one is a token counter you can eyeball

M​iniMax bills these two models on incompatible units, which is the trap in any “which is cheaper” question. H3 is metered in seconds of generated output, by resolution: $0.08 per second at 768P, $0.13 per second at 2K, with input material metered on top — the first five reference images free and $0.04 each after that on the base model, audio input free, reference video billed by its own duration at the output resolution. A 10-second 2K clip is $1.30 of output before any references.

MiniMax-H3-Image is metered in tokens against an image-output rate. On the standard tier the published card reads $2.30 per million input text tokens, $1.50 per million input image tokens and $3.80 per million output image tokens, and the vendor's own worked examples resolve that to:

• 1K (1024×1024) text-to-image — $0.02179 per image.

• 2K (2048×2048) text-to-image — $0.02333 per image.

• 4K (4096×4096) text-to-image — $0.06879 per image, listed at $0.03439 under a limited-time 50% discount.

• 1K edit with one reference — $0.02654; 2K — $0.02658; 4K — $0.07312, discounted to $0.03656.

• 1K edit with three references — $0.03563; 2K — $0.04177; 4K — $0.07914, discounted to $0.03957.

The shape of that ladder is the interesting bit. Sixteen times the pixels costs a little over three times the money, because the meter is tokens and the token count does not scale linearly with area the way intuition suggests. The high-speed variant, MiniMax-H3-Image-Highspeed, takes identical parameters and limits and bills $7.60 per million output image tokens — $0.03529 at 1K, $0.13001 at 4K before the same discount. The discount applies when the delivered image is at least 16,600,000 pixels and runs until 17 November 2026 at 16:00 UTC, which is the only dated commercial fact in the entire release.

Put the two meters next to each other and the fork gets a price. One second of 768P video is $0.08; the cheapest M​iniMax still is about $0.022. A four-second clip — the shortest H3 will make — costs roughly fifteen 1K images. That is not an argument for either model. It is the reason “which is cheaper” has no answer until you say what you are producing.

Where the fork bites in a real deployment

Three asymmetries matter more than the benchmark question.

The first is input handling. H3-Image takes up to 16 reference images in one edit and refers to them positionally as <Picture 1>, <Picture 2> and so on, in upload order; total request body is capped at 64 MiB, and Base64 inflates a file by about a third, so a 16-image request has to use URLs or pre-uploaded file references rather than inline data. H3's reference mode is a different animal — 9 images, 3 videos, 3 audio clips, 12 files, all packed into one sequence with temporal structure. Neither model's reference handling is a superset of the other's.

The second is prompt length. H3-Image's served API accepts up to 32,000 characters, but with prompt enhancement disabled only roughly the first 4,096 tokens take effect; the local SGLang path truncates at 4,096 tokens, full stop. That is not a large budget for a model whose selling point is detailed instruction following, and it is why the companion prompt expander exists at all — MiniMax-H3-Image-PE, a 27B FP4 fine-tune of Qwen3.8-27B that rewrites a short request into a long description plus an aspect-ratio recommendation, released on 8 November and covered by the same community licence. Its editing half shipped as a prompt only; the weights for edit-side expansion are still described as in development.

The third is the licence, and it is the same licence for both models in name only. The MiniMax-H3-Image Community License Agreement requires separate prior written authorisation from M​iniMax if your annualised revenue, or your affiliates', exceeds one million US dollars, or if you provide hosted services to third parties. It requires you to display “MiniMax-H3-Image” prominently on the interface of any commercial product that uses it. It restricts use to the defined Applicable Territory, forbids using outputs to improve any other AI model, and claims no rights over the outputs themselves. The H3 video licence is its own agreement with its own terms. If you are building a product on both, you are holding two licences, not one.

What neither model does

MiniMax-H3-Image cannot make a video. It has no temporal axis, no audio branch, no duration parameter — the audio VAE is simply gone from the architecture, and the output is one latent image per request, up to ten per call through the API. MiniMax-H3 cannot make a 16-megapixel still. It generates 768p by default and reaches 2K through a regeneration pass; asking it for a 4096×4096 single frame is not a supported request. If your pipeline needs a hero still and a five-second loop with synced dialogue from the same visual world, you are calling two models, and on M​iniMax's platform you are paying two different meters.

Both do share one thing that is easy to miss: no negative prompt. H3-Image rejects the field. H3's prompt guidance steers through positive description and through Context-IR rather than through exclusion. Anyone porting a Stable Diffusion workflow into either model loses that lever on day one.

Which one to call

The honest division of labour is by deliverable, not by quality. If the artefact is a still — a product render, a keyframe, a 16-megapixel plate with legible text in it — MiniMax-H3-Image is the model, and the six-step schedule means you can afford to iterate on it. If the artefact moves, H3 is the model, and the audio branch is the reason to prefer it over stitching stills into a sequence.

What a routing layer changes here is narrower than it sounds, and worth stating plainly: OrcaRouter does not serve MiniMax-H3-Image. The model reached public weights on 10 November and has no route on our catalogue, so if you want it today you get it from M​iniMax's own API or run the 68.65 GB checkout yourself. What we do carry is minimax/minimax-h3, the video model, billed per second at list price with no markup — and the useful property for a two-model pipeline is that the same key that calls H3 also reaches the other 200-odd models on the platform.

The thing to watch is not whether H3-Image is good. It is whether M​iniMax says anything about it. A model that ships with weights, a licence, a pricing card and four framework integrations, but no blog post and no benchmark, is a model whose claims are still entirely the vendor's. The first independent placement on an image board, or the first head-to-head against a closed model at the same resolution, will tell you more than the showcase grid does. Until then the repository is the record, and the repository says: this is H3 with the sound switched off.