MiniMax H3 Max: The Spec, the Rate Card, and the Tier Decision

This page is not a launch story. MiniMax H3 Max dates from August 2026, and the model it is built on, MiniMax H3, is older still — MiniMax shipped that one on 31 July 2026, and its weights are public. What makes this page worth writing today is not freshness: it is that the bare name “minimax h3 max” was typed 6,504 times in the 28 days to 25 September, and the best position we hold on that exact query is 8.4 — near enough to be seen, far enough that no page of ours is the answer a searcher is looking for. So this is the reference page: what MiniMax H3 Max is, where it sits against MiniMax H3, what MiniMax's own documentation says it costs and accepts, and what an independent board scores it at — every figure read from a primary source on 28 September 2026.

The short version of the tier question, because it is the one readers actually ask: MiniMax H3 Max is the fast tier, not the big one. MiniMax's developer documentation describes MiniMax H3 Max as a video generation model post-trained on MiniMax H3 and released jointly by MiniMax and a partner lab, optimized for high-speed generation. It outputs 480P or 768P and 5-to-15-second clips. MiniMax H3 outputs 768P or 2K and 4-to-15-second clips, is the one with open weights, and is the one that can edit video rather than only generate it. “Max” here buys you speed and a lower entry price, and it costs you the resolution ladder, the 2K upscaler, and the open release.

What MiniMax H3 Max is

MiniMax's video generation guide opens by stating that the company currently offers two video generation models, MiniMax H3 and MiniMax H3 Max, and it is the clearest single source for the split. MiniMax H3 is described as an open, general-purpose multimodal video model that understands text, image, video and audio inputs in a unified way and supports video generation, reference-based creation and video editing. MiniMax H3 Max is described as a video generation model post-trained on MiniMax H3 and optimized for high-speed generation, delivering “mainstream 480P and 768P output” and generating faster than MiniMax H3.

Every one of those phrases is the vendor's own, and the speed advantage in particular is a vendor claim: MiniMax publishes no latency or throughput figure for H3 Max anywhere in its developer documentation, so “faster” is documented qualitatively and not quantified. What is documented precisely is the envelope, and the envelope is where the two models actually diverge.

One thing worth noting for anyone tracking the family: MiniMax's release-notes page, which is the canonical dated index of the company's model launches, carries an entry for MiniMax H3 dated 31 July 2026 and no entry at all for MiniMax H3 Max. The model exists in the developer documentation, the API reference and the pricing card, and it is identifiable as an H3 post-train, but MiniMax never gave it a launch-post date of its own. Artificial Analysis dates it to August 2026 on its boards.

The tier question: MiniMax H3 vs MiniMax H3 Max

Almost all of the loose search traffic around this model is the tier comparison — “minimax h3 vs minimax h3 max” is the single largest of the secondary queries around the name, and the pricing and local-deployment questions behind it are the next two. Here is the comparison on the dimensions MiniMax's own spec sheet defines, with both sides stated:

• Output resolution — MiniMax H3: 768P or 2K. MiniMax H3 Max: 480P or 768P, with 768P as the default and 2K not supported.

• Clip length — MiniMax H3: 4 to 15 seconds, integer values only. MiniMax H3 Max: 5 to 15 seconds, integer values only; 4 seconds is rejected.

• Modes — both support text-to-video, image-to-video with first and/or last frame, and reference generation from reference images, video or audio. MiniMax H3 additionally supports video editing, and the 2K-regeneration and context-IR endpoints are MiniMax H3 only.

• Speed — MiniMax H3 Max is documented as the faster of the two; MiniMax H3 is not. No figure is published for either.

• Open weights — MiniMax H3 has a public checkpoint and a self-hosting path. MiniMax H3 Max has no published weights. Artificial Analysis's own FAQ names MiniMax H3, not H3 Max, as the open-weights leader on both video boards.

• Entry price at 768P — both list at $0.08 per second. The cheaper tier is H3 Max at 480P, $0.05 per second.

So the tier decision is not “which is better” — it is whether you need 2K or open weights. If you need either, MiniMax H3 Max is not a candidate no matter how much faster it is. If you are generating high volumes of 768P or 480P short clips and discarding most of them, H3 Max is the tier that makes the discard rate affordable.

Inputs and outputs, from MiniMax's spec sheet

This family is not text-only, and the input surface is the most under-reported part of it. MiniMax's video generation guide gives one shared input-requirements table for both models, which means what H3 Max accepts for references is what H3 accepts:

• Modalities in — text, images, video clips and audio clips, all as reference material for a generated video. Output is video, with the duration and resolution you declare.

• First and last frame — zero, one or two images, each between 256 and 5,760 pixels on a side, with an aspect ratio between 2:5 and 5:2. Zero images is text-to-video.

• References — up to 9 images; up to 3 video clips of 2 to 15 seconds each, 15 seconds total; up to 3 audio clips of 2 to 15 seconds each, 15 seconds total. Mixed input is capped at 12 files. Image-to-video and reference-to-video are mutually exclusive in one request.

• Formats and caps — video H.264/AVC or H.265/HEVC with AAC or MP3 audio; images JPG, JPEG, PNG, WEBP, HEIC or HEIF; audio WAV or MP3. Video up to 50 MB per file, images 30 MB, audio 15 MB, request body 64 MB, prompt up to 7,000 characters.

• Aspect ratio — adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Text-to-video requires an explicit ratio and cannot use adaptive; image-to-video always resolves to adaptive from the input frame.

• H3 Max only — an extra.prompt_expansion_mode field with three settings: disabled, balanced (the default) and quality. No equivalent is documented for MiniMax H3.

The workflow is asynchronous for both: you POST to the video generation endpoint with the model name in the model field, receive a task ID, poll it, and download from the returned URL. Both models are the same endpoint and the same request shape, which is the practical point — the tier is a field value, not an integration.

The rate card

MiniMax's pay-as-you-go card, read on 28 September 2026, lists output pricing per second of generated video. MiniMax H3 Max is $0.05 per second at 480P and $0.08 per second at 768P. MiniMax H3 is $0.08 per second at 768P and $0.13 per second at 2K. These are vendor list prices and they are the numbers that apply today.

Input material is billed on top, and the two tiers bill it differently — this is the part that catches people out. For MiniMax H3 Max, audio references are free, the first two images are free and each additional image is $0.074, and a reference video is billed by its input duration and your output resolution: $0.0553 per second at 480P, $0.143 per second at 768P. For MiniMax H3, the free image allowance is five rather than two, images beyond that are $0.04 each, and a reference video bills at the output rate — $0.08 per second at 768P, $0.13 at 2K. So a reference-heavy workflow is not simply cheaper on the Max tier: it gives you two free images instead of five, and it charges more per second of reference video at 768P than H3 does.

Two further billing facts from the same card. First, 2K on the H3 family is a separate paid step: the regeneration endpoint takes a 768P result to 2K at $0.05 per second of regenerated output, and the context-IR endpoint that produces an enriched structured prompt out of your multimodal inputs bills $0.90 per million input tokens and $3.60 per million output tokens. Second, and more consequential for budgeting, MiniMax's prepaid video packages cover the older Hailuo video models, and the packages page states that MiniMax H3 is not supported by them and points H3 users to the pay-as-you-go API; MiniMax H3 Max appears only on the pay-as-you-go card.

One number needs a flag. Artificial Analysis tracks MiniMax H3 Max at $2.40 per minute on its video boards, and its footnote defines that column as the cost of generating one minute of 1080p video on the creator's API at default settings. MiniMax's own card says $0.08 per second at 768P, which is $4.80 per minute. The two do not agree, and 1080p is a resolution this model does not produce, so the independent column is a normalised comparison across models rather than a quote for H3 Max. Where they conflict, the vendor card is the figure that applies to your bill; we found no current MiniMax page documenting a discount that would close the gap.

The independent read

There is no vendor benchmark for MiniMax H3 Max — MiniMax publishes no score for it, and the release-notes page does not even list it. The only independent measurement is Artificial Analysis's video arena, which is blind human-preference voting, not a benchmark suite, and it is the one number here that is not the vendor talking about itself.

Captured on 28 September 2026, MiniMax H3 Max sits at #1 on the with-audio image-to-video board with Elo 1,194 across 5,765 votes, ahead of MiniMax H3 at 1,181, Gemini Omni Flash at 1,178, Dreamina Seedance 2.0 720p at 1,176 and HiDream-O1-Video at 1,174. On the with-audio text-to-video board it is #3 with Elo 1,227 across 5,689 votes, inside a four-model tie range at the top: Gemini Omni Flash leads at 1,233 and Wan 3.0 is at 1,229. Those figures have moved since August — the same boards showed 1,204 and 1,235 at the time of the model's first appearance — which is what an arena does as votes accumulate, and it is the reason to read these as a current snapshot rather than a settled ranking.

What the arena does not tell you is whether the post-train beat its parent on quality in any controlled sense. MiniMax H3 and MiniMax H3 Max have not been benchmarked against each other at matched effort on a matched harness — nobody runs that comparison for video models, and the two are being voted on as separate entries with separate vote counts. The honest statement is the narrow one: on this board, on this date, H3 Max is ahead of H3 by 13 Elo points, and Artificial Analysis's own FAQ names H3 Max the leader among image-to-video models with audio output.

What MiniMax does not document

A reference page that only repeats the vendor's marketing is not much of a reference page, so here is what could not be established. There is no published latency, throughput or real-time factor for MiniMax H3 Max — the speed advantage is documented as a direction, not a number. There is no H3 Max entry in MiniMax's release notes and no dated launch post. There are no published weights: the MiniMaxAI organisation on Hugging Face lists MiniMax H3 and no H3 Max, and the self-hosting documentation covers ComfyUI, SGLang and vLLM-Omni routes for MiniMax H3 only. There is no 2K, no video-editing mode and no context-IR or regeneration endpoint on the Max tier. And the open MiniMax H3 it derives from carries a community licence whose revision of 26 August 2026 excludes the United States, the European Union, the United Kingdom and South Korea, and requires a separate licence for those territories — a constraint that applies to MiniMax H3, since H3 Max has no weights to license at all.

Calling either tier

MiniMax's own route for both models is the pay-as-you-go API on its platform, so your integration cost is one endpoint with the model name switched. If you want the family behind a single key, MiniMax H3 is in our catalogue on OrcaRouter — the open, 2K-capable, editable one — at the provider's list price of $0.08 per second at 768P and $0.13 per second at 2K, with 0% markup, so any change to that card reaches you the same day rather than at a marked-up rate, and with automatic failover so a single upstream wobble does not take your generation queue down with it. That matters more than usual in video, where a request is a long-running job rather than a call you can quietly retry.

Which tier to take comes down to one question, and it is not about Elo. If your output has to be 2K, or has to be self-hosted, or has to edit existing footage, MiniMax H3 is the only option and the speed gap is the price of those capabilities. If your output is 480P or 768P social, ad or e-commerce volume, where you generate more clips than you keep, MiniMax H3 Max is the cheaper tier to throw work away on, and the arena currently rates it the better of the two. What neither the board nor the spec sheet settles is the price question: an independent tracker showing $2.40 per minute against a vendor card charging $0.08 per second is a gap somebody will eventually explain or correct, and it is the one number on this page to re-read before you commit a budget to it.

Every routed model is charged at the provider's list price, passed through with 0% markup, so 200+ models behind one API — and a vendor price change lands the same day rather than at the next contract renewal.