MiniMax H3: the top-ranked model you can also download

What MiniMax H3 is

MiniMax H3 is the video generation model that replaced the Hailuo line during 2026, succeeding Hailuo 2.3. It is the current model inside MiniMax's own consumer app and it is sold on the company's pay-as-you-go API under two identifiers: MiniMax-H3 and MiniMax-H3-Max, the latter developed jointly with fal.ai.

The naming takes a sentence to untangle, because two systems are running at once:

• MiniMax is the company and the developer-facing model name. The API identifiers are MiniMax-H3 and MiniMax-H3-Max.
• Hailuo is the consumer product name — the app — and also the name of the previous model generation, Hailuo 2.3, which MiniMax now files under Legacy Models.

So "Hailuo H3" is reasonable shorthand for using H3 through the Hailuo app, but the model is named MiniMax H3, and searching for the old name will surface a generation that has already been superseded.

What makes this model worth a page of its own is a combination that no competitor currently offers: it is ranked first on a major blind leaderboard and its weights are downloadable. Gemini Omni, Seedance and Kling are all closed. H3 is not.

H3 vs H3 Max — and why the cheaper one is not always cheaper

H3 Max is a speed-optimised variant derived from H3 through additional training, aimed at faster generation and better prompt adherence. The published differences are narrower than the naming suggests, and one of them moves in the wrong direction.

Read the resolution and rate rows together and the pitch is clear: H3 Max buys you a cheap 480p tier, and charges the same as H3 at 768p. It is not a discount on the model you already had — it is a new floor underneath it.

The row that catches people is the image inputs. H3 Max charges nearly twice as much per additional image and allows three fewer for free. MiniMax works through an example that reverses the intuition: a 15-second 768p request with four reference images and a six-second reference video comes to roughly $2.21 on H3 Max against about $1.68 on H3. The "cheaper" model costs a third more, because the request is input-heavy. If your workflow is reference-driven, H3 is the cheaper route at every resolution they share.

What it costs

MiniMax bills per second of finished video on its pay-as-you-go API. There is no credit system to decode at this level, which makes it one of the easier models to budget.

Inputs are billed separately and are where the surprises live. Reference video is charged by input duration at the output resolution's rate — $0.0553 per second for 480p output, $0.143 for 768p — so a six-second reference clip is a real line item rather than a rounding error. Audio references are free on both models.

For context, the previous generation is still listed and is not obviously worse value for short clips: Hailuo 2.3 runs $0.28 for a six-second 768p clip and $0.49 at 1080p, and Hailuo 2.3 Fast is image-to-video only. MiniMax has moved those to Legacy Models, so new workflows are expected to evaluate H3 and H3 Max first.

What's built on H3 — and why it can beat H3's own price

H3 is not only sold by MiniMax. In late September 2026, Pruna AI launched P-Video-2-Pro, describing it in its own launch post as "based on MiniMax H3" — a derivative rather than a wrapper, tuned for generation speed and distributed through partner platforms rather than through MiniMax.

What it does differently is narrow and specific:

• Input: text or a first-frame image, with optional last-frame conditioning — the same frame-steering logic as H3, and consistent with what this model family is best at.
• Speed: Pruna claims roughly 2.0 seconds to produce a 5-second 480p clip and about 4.3 seconds at 768p in its Speed mode. That is faster than real time, which is the point of the model.
• Output: 24 fps with generated audio included. Notably, audio input is not exposed — so the free audio references that MiniMax offers on H3 have no equivalent here.
• Length: 5 to 15 seconds, 480p or 768p, with no 2K tier. Billing is measured from the finished video, and an omitted mode is billed as Speed.

Then there is the price, which is the part worth sitting with. Pruna publishes three generation recipes — cost, speed and quality — and the rates look like this against MiniMax's own:

At 768p the derivative undercuts the original at every setting, and its cheapest mode is roughly a third of MiniMax's own rate. At 480p the gap is wider still: P-Video-2-Pro starts at $0.01 per second in cost mode, against $0.05 for H3 Max at the same resolution.

Two things follow from this that are worth knowing before you choose where to generate. A "MiniMax H3" price is now a platform price, not a model price — which means any comparison table that quotes one number for H3 is quoting one vendor's opinion. And the H3 family's competitive claim is increasingly about serving cost rather than output quality: with MiniMax's own open weights, a speed-tuned commercial derivative and nine partners all in play, the question has shifted from whether the model is good to who can run it cheapest.

The open weights — what you actually get

MiniMax published H3's weights on Hugging Face under the MiniMax H3 Community License. This is the part that separates H3 from every model it competes with, and also the part where the practical detail is usually omitted.

Two licence conditions are easy to miss and both have teeth:

• A commercial product with more than $20 million in annual revenue requires written permission from MiniMax. This is the term that makes the open release unsuitable, on its own, for larger companies without a conversation.
• Commercial products must display "MiniMax H3" in their interface. Not in your documentation — in the product. That is a user-visible attribution requirement, and it needs to be designed in rather than retrofitted.

Neither condition applies if you use the hosted API instead, which is the trade: the API costs money per second but carries no attribution or revenue threshold.

How it ranks — and what the pattern tells you

Arena AI runs blind pairwise comparisons with human voters and scores models with an Elo system. MiniMax H3's placement across three boards is unusually informative, because it is not uniform.

First on image-to-video, third on editing, eighth on text-to-video is not noise — it is a profile. This model is strongest when it has an image to work from: animating a still, holding a character that has been supplied, keeping a scene consistent across an edit. It is comparatively weaker when asked to invent a scene from a sentence. Everything else about the model is consistent with that, including the fact that MiniMax's own pricing spends its detail on how images and reference video are billed.

If you are choosing between H3 and Gemini Omni, that distinction is the deciding one. Gemini Omni's advantage is conversational editing and text rendering on a prompt-first workflow; H3's is fidelity when you already know what the first frame looks like. Pick the model that matches where your input comes from.

Is Hailuo free? Mostly no

This deserves a direct answer because the old assumption no longer holds. As of September 2026 the Hailuo free plan carries no monthly credits. What it offers is one-time trial credits for MiniMax's own models — an evaluation allowance rather than a renewable one. You can test with it; you cannot work with it.

There have been periodic promotions offering large credit bundles, including for installing other MiniMax apps, but those are promotions and should not be part of a plan. The two routes that are genuinely free are the open weights, which cost nothing to download and require four GPUs, and whatever free allowance an individual third-party host chooses to offer on its own terms — which is not MiniMax's price and can change without notice.

Getting a usable clip

1. Decide which model by your input, not your budget.

   Reference-heavy work is cheaper on H3 because additional images cost $0.04 rather than $0.074 and the first five are free. If your request carries audio reference or a reference video, run the arithmetic first — input-heavy jobs can cost a third more on H3 Max despite the lower headline rate.

2. Use 480p to block out, only if 480p is acceptable output.

   The $0.05 tier on H3 Max is the cheapest route into this family, but there is no path from 480p up to 2K — that tier simply does not exist on Max. If the final has to be 2K, you are on H3 and the 480p saving is not available to you. Plan the resolution before the rate.

3. Write it as a shot, not a concept.

   Name the subject, the action, the place, the camera move and the light. "Espresso pours into a glass cup in slow motion, morning window light, shallow depth of field" gives a model far more to work with than "coffee video" — and this model's weaker text-to-video placement makes prompt discipline matter more, not less.

4. Steer with frames, not words.

   H3 Max accepts a first frame, a last frame, or both. Given that image-to-video is where this model ranks first, supplying the frames is playing to its strength; describing them in prose is playing to its weakest board.

5. Keep clips inside the duration you are billed for.

   Billing is per output second, so an over-long generation is not just slower — it is directly more expensive. Decide the duration before you generate rather than trimming afterwards, since there is nothing to refund.

6. Check the licence before you build on the weights.

   If you plan to ship a commercial product on the open release, two things need handling early: the $20 million revenue threshold requires written permission, and the interface must carry the "MiniMax H3" attribution. Both are design decisions, not paperwork.

Frequently asked questions

What is MiniMax H3?

The video model that replaced the Hailuo line in mid-2026, succeeding Hailuo 2.3. It is sold on MiniMax's pay-as-you-go API at $0.08 per second of 768p output and $0.13 per second at 2K, and it is one of the very few frontier video models whose weights have been published for download. On Arena AI's blind image-to-video leaderboard it held first place at an Elo of 1495 across 5,711 votes.

What is the difference between MiniMax H3 and H3 Max?

H3 Max is a speed-optimised variant derived from H3 by additional training, released jointly with fal.ai. It adds a cheap 480p tier at $0.05 per second but has no 2K option, and it matches H3's price at 768p rather than beating it. Its image inputs are also more expensive — two free then $0.074 each, against five free then $0.04 on H3 — so input-heavy requests can cost more on H3 Max than on H3.

How much does MiniMax H3 cost?

H3 is $0.08 per second at 768p and $0.13 at 2K; H3 Max is $0.05 at 480p and $0.08 at 768p. A ten-second clip is therefore $0.50, $0.80 or $1.30 depending on the route. Inputs bill separately — reference video at $0.0553 per input second for 480p output or $0.143 for 768p — while audio references are free on both models.

Can I run MiniMax H3 locally?

Yes, within limits. The weights are on Hugging Face under the MiniMax H3 Community License. The main transformer is 33B parameters and the pipeline also runs the full Qwen3-VL-32B model to read prompts and images, with roughly 66 GB per checkpoint, and MiniMax's own serving example uses four GPUs. The open release is H3-Base at 768p only — the 2K regenerate step and the sparse attention optimisation are not open-sourced.

What are the licence terms for the MiniMax H3 open weights?

Two conditions matter. A commercial product with annual revenue above $20 million needs written permission from MiniMax, and commercial products must display "MiniMax H3" in their interface. The first rules out large companies without a conversation; the second is user-visible attribution that has to be designed in. Neither applies when you use the hosted API.

Is Hailuo AI free?

Largely not. As of September 2026 the free plan carries no monthly credits — only one-time trial credits for MiniMax's own models, which is an evaluation allowance rather than a renewable one. Promotions have offered large credit bundles for installing other MiniMax apps, but those are not a plan. The genuinely free routes are the open weights, which need serious hardware, and whatever trial an individual third-party host offers on its own terms.

How is MiniMax H3 ranked?

First on Arena AI's image-to-video board (Elo 1495, 5,711 votes), third on video editing (1392) and around eighth on text-to-video (1460). That profile is the useful part: the model is strongest when it is given an image to work from and comparatively weaker when asked to invent a scene from a prompt.

What resolutions and clip lengths does MiniMax H3 support?

H3 Max produces 5 to 15 second clips at 480p or 768p, in integer durations, with aspect ratios covering 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, steerable with a first frame, last frame or both. H3 adds a 2K tier. Note that at least one reseller describes H3 Max as not supporting reference-image, reference-video or reference-audio conditioning, which conflicts with MiniMax's own image-input billing — verify on the platform you actually buy from.

Is MiniMax H3 still called Hailuo?

Both names are current but they describe different layers. MiniMax is the company and the developer-facing model name, with API identifiers MiniMax-H3 and MiniMax-H3-Max. Hailuo is the consumer product name and the previous model generation, Hailuo 2.3, which now sits under Legacy Models. "Hailuo H3" is fair shorthand for using H3 through the Hailuo app, but the model is named MiniMax H3.

Should I use MiniMax H3 or a hosted aggregator?

MiniMax's own API gives per-second billing, the full resolution range and the official capability list, but you build the editing, voiceover and captions yourself. Aggregators such as Higgsfield wrap several vendors behind one subscription and one credit balance — convenient for comparison, but the host sets its own prices and free allowances, they are not MiniMax prices, and a credit is not a unit of video. The open weights are a third route for teams with GPU capacity.

Is there a cheaper way to use MiniMax H3 than MiniMax's own rate?

Yes — and this is the most useful thing on this page. In late September 2026 Pruna AI launched P-Video-2-Pro, which it describes as based on MiniMax H3. At 768p it undercuts H3 direct at all three of its settings: $0.075 per second in quality mode, $0.035 in speed mode and $0.025 in cost mode, against MiniMax's $0.08 for the same resolution. At 480p it starts at $0.01 against $0.05 for H3 Max. The caveats matter, though: it is a different, speed-tuned model rather than the same one; its benchmark results were still pending at launch, so the quality claim is the vendor's; and it ships through nine inference partners that each set their own rate. The upshot is that a MiniMax H3 price is now a platform price, not a model price.

Also on this site

• Gemini Omni — the model H3 beats on image-to-video and trails on text-to-video. Google's any-to-any video model, its free YouTube route and its real per-second cost.
• Seedance 2.5 — ByteDance's closed video model, and the four specifications no engineering source confirms. The contrast with H3's open release is the point.
• Kling 4.0 — the competitor promising 30-second single generations. What Kuaishou has actually shipped versus what it has only announced.
• Nano Banana 2.1 — the image side of the same problem: what Google halved the price on, and where Pro still beats it.
• GPT Image 2.5 — OpenAI's two-models-one-price release, and the transparent-background capability almost nobody mentions.
• Higgsfield AI review — the aggregator route: what each plan grants, what a clip really costs, and which surfaces bill even on an unlimited model.
• Grok Imagine — the closest thing to a price rival, and the one with a public per-second rate on its own site. The free tier came back in October 2026 with a hard cap of one 480p clip a day.