MiniMax H3 vs MiniMax H3 Max: 2K Omni-Ref or Faster 768p Drafts?

TLDR

MiniMax H3 is MiniMax’s own omni-modal video model, launched 31 July 2026: unified text / image / video / audio context, clips up to 15 seconds, a 2K delivery path, and native stereo in the same pass (official H3 post; Reuters). MiniMax H3 Max is a hosted, post-trained variant of those H3 weights, public from 26 August 2026. It is not a MiniMax-branded H3 upgrade: it generates 480p or 768p (not 2K), text-to-video and image-to-video (optional last frame), 5–15 seconds, with native audio. Reach for H3 Max when you want a fast 768p draft; stay on H3 on this site for 2K text-to-video or image-to-video (optional last frame). MiniMax’s H3 spec still includes omni-reference — up to nine images, three videos, three audio clips — but this site’s generator cannot attach those extra video or audio files today (H3 post; open-source notes).

Key Takeaways

• Two products, one family. H3 is MiniMax’s July model (H3 post). H3 Max is a later hosted post-train of the H3 open weights, announced in third-party rankings on 26 August 2026 (Artificial Analysis; Design Arena).
• Resolution is the first split. H3’s public story is 2K (the open H3-Base still renders a 768 short edge; 2K is the regenerate stage) (open-source notes). H3 Max stops at 768p (with a 480p draft tier).
• References are the second split. MiniMax’s H3 spec binds up to nine images, three videos, and three audio clips in one request (H3 post; open-source notes). H3 Max at launch is text-to-video plus image-to-video, with an optional end frame — not that full stack. This site’s H3 generator is also text-to-video and image-to-video only (2K, optional last frame). It does not take mixed video or audio references today.
• Speed is why Max exists. Design Arena’s 26 August post reported image-to-video ~6.4s and text-to-video ~4.7s, and described Max as far faster than base H3 on that arena (Design Arena). Treat those as arena wall times, not a service-level promise.
• Rankings, snapshot 27 August 2026. On Artificial Analysis image-to-video with audio, H3 Max is #1 (Elo 1,204, ±10, 5,123 samples) and open-weights H3 is #3 (Elo 1,184) (I2V board). On text-to-video with audio, H3 Max is #3 (Elo 1,235) and H3 is #4 (Elo 1,225) — the top three sit within a few points (T2V board).
• Weights. H3 checkpoints are public (Hugging Face MiniMax-H3). H3 Max weights are not public as of this writing. Independent write-ups noted an intent to release them; that is not the same as a download (Artificial Analysis).

Head-to-head

| Dimension | MiniMax H3 | MiniMax H3 Max | | --- | --- | --- | | What it is | MiniMax’s omni-modal video model (H3 post) | Hosted, post-trained H3 variant for speed and prompt follow | | Public window | 31 July 2026 (Reuters) | 26 August 2026 third-party boards (AA) | | Resolution | 2K path; open base default short edge 768 (open-source notes) | 480p or 768p — no 2K tier | | Duration | 4–15s in MiniMax’s open-source spec (open-source notes) | 5–15s | | Audio | Native stereo with the picture (H3 post) | Native audio with the picture | | Inputs | Text, images, video, audio; omni-reference (H3 post) | Text-to-video and image-to-video; optional last frame | | Speed signal | Slower hosted path on the same arenas | Design Arena: I2V ~6.4s, T2V ~4.7s (post) | | I2V with audio (27 Aug 2026) | Elo 1,184, open-weights lead (AA I2V) | Elo 1,204, #1 on that board | | T2V with audio (27 Aug 2026) | Elo 1,225, #4 (AA T2V) | Elo 1,235, #3 | | Weights | Public H3-Base (Hugging Face) | Not public | | First try | H3 video generator — 2K text or still | H3 Max video generator |

Read the table once: if the brief is iterate 768p text or stills until the motion holds, H3 Max. If the brief needs 2K text-to-video or image-to-video, use H3 on this site. MiniMax’s mixed image / video / audio references are an official H3 spec; this generator cannot take those extra files.

Family / timeline

Keep three names apart.

MiniMax H3 (also listed as Hailuo 3 / Hailuo 03 on some surfaces) is the July 2026 video model in the Hailuo line: Hailuo 01 → Hailuo 02 / 2.3 → H3 (H3 post). It is not MiniMax M3, the June language / agent model (M3 page).

MiniMax H3 Max is not “H3 but official 2K.” It is a later hosted post-train of the H3 weights. MiniMax’s own July and August materials describe H3 and the open-weight drop; they do not announce a MiniMax product named H3 Max. Third-party boards and the 26 August ranking posts are what made the Max name public (Artificial Analysis; Design Arena).

The open-weight split still matters for H3: MiniMax published H3-Base (768p-class generate + stereo) and kept H3-Context-IR and H3-Regenerate-2K as hosted stages (open-source notes). Anything post-trained only from the public base inherits that 768p ceiling unless a separate 2K stage is attached. That is the mechanical reason Max does not replace H3’s 2K story.

For the original H3 launch map, see MiniMax H3 Launch: 2K AI Video With Native Stereo Audio (July 31, 2026).

What H3 looks like in motion

This clip is a community H3 sample re-hosted on this site’s CDN. It shows H3’s commercial density. We do not have a same-prompt H3 vs H3 Max pair here — this is not a lab A/B.

Scenario winners

| If your bottleneck is… | Prefer | Why | | --- | --- | --- | | Many 5–10s drafts in one sitting | H3 Max | Arena speed is the published reason Max exists (Design Arena) | | A still that must keep identity while it moves | H3 Max image-to-video | Start from a frame on the H3 Max image page, then generate | | 2K delivery language, logos, fine type | H3 | 2K is H3’s path; Max has no 2K tier (H3 post; open-source notes) | | Character + motion + voice from mixed files | Official H3 spec — not this generator | MiniMax documents omni-reference for H3 (H3 post). This site cannot attach Video 1 / Audio 1 today. | | First-and-last-frame lock | Either, with a caveat | H3 first/last-frame is a core H3 mode; Max image-to-video can take an optional end frame — still not full omni-ref | | Self-host / fine-tune the open checkpoint | H3 | Public weights are H3-Base (Hugging Face) |

Practical split: draft motion on MiniMax H3 Max until the camera and the beat hold. Promote the keeper to MiniMax H3 when you want the 2K pass of the same text or still. Mixed video/audio references stay an official H3 capability, not something this generator can take.

What We Know vs. What We Don’t

| We know (sourced) | We don’t know / won’t claim | | --- | --- | | H3 launched 31 July 2026: omni-modal, ≤15s, 2K story, native stereo, then open weights (H3 post; open-source notes; Reuters) | That MiniMax itself shipped a product named H3 Max | | H3 Max is a hosted post-train of H3, public on ranking boards 26 August 2026, 5–15s, up to 768p, T2V + I2V (AA) | That Max can output 2K, omni-reference, or instruction-based edit today | | This site’s H3 generator: 2K text-to-video and image-to-video, optional last frame | That you can upload reference video or audio on this generator today | | AA I2V-with-audio #1 for Max (Elo 1,204) and #3 for H3 (Elo 1,184) as of 27 August 2026 (I2V board) | That Elo will stay put next week | | Design Arena posted I2V ~6.4s / T2V ~4.7s for Max (Design Arena) | Queue time on a busy day; “always faster than playback” as an SLA | | H3 weights are downloadable; Max weights are not, as of this writing (Hugging Face) | A date for Max weights |

How to evaluate both

Use the same Prompts A and B on both generators. Do not invent a new brief halfway. Prompt C is spec-only.

1. Open MiniMax H3 Max and run Prompt A at 5s / 768p.
2. Open MiniMax H3 on this site and run Prompt A again.
3. Judge subject lock, camera follow, audio fit, and whether on-screen type (if any) stays readable.
4. If you already have a still, run Prompt B as image-to-video on Max, then the same still on this site’s H3 generator.
5. Prompt C is a read-only spec check for MiniMax’s official H3 omni-reference. Do not paste it into this site’s generator — that UI cannot take Video 1 or Audio 1.

This is not a lab A/B. It is the shortest honest test: one sentence, two generators, one decision.

Prompt A — product hero (16:9, both models)

5s luxury product film, 16:9, photoreal. One continuous push-in on a brushed metal watch on marble, window light crawling across the crystal, shallow depth of field. Native stereo: quiet room tone and a single tick. No extra logos. Cinematic grade.

Prompt B — still to motion (image-to-video)

Keep the uploaded still as the first frame. Slow 10s orbit around the subject, small amplitude, no identity drift, no new wardrobe, no extra people. Native stereo: soft room tone only. 16:9 if the still allows it.

Prompt C — official H3 omni-reference (not runnable here)

This prompt is for reading MiniMax’s H3 spec. It is not a step on this site.

Match character identity from Image 1, camera language from Video 1, and vocal energy from Audio 1. 12s fashion lookbook walk through a rain-wet night street, neon reflections, confident pace, native stereo footsteps and city hush.

FAQ

Is MiniMax H3 Max just a faster MiniMax H3?

No. Max is a hosted post-train of H3, tuned for speed and prompt follow at 768p, with text-to-video and image-to-video. MiniMax’s H3 spec still covers 2K and omni-reference (H3 post; AA). On this site you can run H3 as 2K text-to-video and image-to-video only.

Did MiniMax release H3 Max?

Public ranking posts on 26 August 2026 introduced H3 Max as a post-trained H3 variant (Artificial Analysis; Design Arena). MiniMax’s own H3 launch and open-weight notes describe H3, not a MiniMax product named Max (H3 post; open-source notes).

Can MiniMax H3 Max do 2K?

Not on the path we can stand behind. Max is 480p / 768p. For 2K, use MiniMax H3 (open-source notes).

Does H3 Max generate audio?

Yes. Public Max descriptions include native audio with the picture, in the same family as H3’s native stereo (AA; H3 post).

How long can a clip be?

H3’s open-source spec is 4–15 seconds. H3 Max on the generator we point to is 5–15 seconds. Plan one beat per generate (open-source notes).

Is H3 Max #1 at everything?

No. As of 27 August 2026 it leads Artificial Analysis image-to-video with audio. On text-to-video with audio it is #3, inside a tight top group (I2V; T2V). Rankings move.

Can I download H3 Max weights?

Not as of this writing. H3-Base is on Hugging Face. Treat any “Max will be open” line as intent, not a file.

Where do I try each one?

H3 Max drafts: minimaxh3max.com video generator. H3 2K text or still (optional last frame): H3 video generator on this site. Have a still first? Start on the H3 Max image page. Mixed image + video + audio references are in MiniMax’s H3 spec, not on this generator today.

Should I use both?

Yes, in sequence: Max to find the motion, this site’s H3 for the 2K text or still pass. MiniMax’s mixed-file omni-reference is outside this generator.

Is Hailuo 03 the same as H3 Max?

No. Hailuo 03 / Hailuo 3.0 usually means MiniMax H3. H3 Max is the later hosted post-train (H3 post).

About Casey Ren

Casey Ren is an AI Video Model Analyst who tracks MiniMax / Hailuo video releases and turns public launch claims into clear try paths on MiniMax H3. Facts in this post follow linked primary sources — not private lab access.