Wan 3.0 vs MiniMax H3: The Open Weights One Has a Catch

Here is the version of this comparison that most coverage gets backwards. MiniMax H3 is the open-weight model — you can download it, inspect it, run it on your own hardware, and it is the only one of these two you can call through a routing API today. Wan 3.0 has no published weights at all, continuing a retreat from openness that started well before this release. So far, so familiar. The part that gets left out is the licence: MiniMax H3's weights come under a community licence that explicitly excludes the United States, the United Kingdom, the European Union and South Korea, where use requires an application to MiniMax. "Open weights" and "open licence" are not the same claim, and on this pair the difference decides more than the spec sheet does.

Both models are current and both are strong. Wan 3.0, from Tongyi Lab, went into public beta on 6 August 2026 and launched officially on 24 August; it generates up to 30 seconds in a single pass at up to 1080p. MiniMax H3, from MiniMax, opened its weights in early August 2026 and generates 4 to 15 seconds at 768p, or 2K through an API-only upscaling module, with native stereo audio. They are the two halves of the same argument about what a video model should be.

The licence question, before anything else

Start here, because if you are in one of the excluded territories it makes the rest of the comparison academic.

• MiniMax H3 — H3-Base weights are downloadable under MiniMax's community licence, not Apache or MIT. The licence excludes the US, UK, EU and South Korea. Users there apply to MiniMax for permission. What is open is the base model only: the Context-IR prompt-expansion module and the 2K regeneration module stay API-side and were never part of the release.

• Wan 3.0 — no weights, no repository, no licence published. Alibaba's launch material describes a metered, closed API and says nothing about opening anything.

That is a genuinely different kind of closed. MiniMax shipped something with terms you may be able to meet. Alibaba shipped nothing to meet.

What each model actually is

Strip the framing and the two are aimed at different jobs.

• Maximum length — Wan 3.0 runs 2 to 30 seconds in one pass. MiniMax H3 runs 4 to 15 seconds. This is the single largest functional difference between them and it is not closable with stitching, because the seam is what you were trying to avoid.

• Resolution — Wan 3.0 outputs 480p, 720p or 1080p, defaulting to 1080p. MiniMax H3 outputs 768p natively, or 2K via a regeneration module that is API-only and therefore unavailable to anyone self-hosting the weights.

• Frame rate and audio — Both generate 24 fps at 768p. MiniMax H3 publishes a clear spec for its native audio: 32 kHz stereo, generated in the same pass as the video. Wan 3.0's launch material describes native audiovisual output and synchronized micro-expressions; one comparison site still lists Wan 3.0's native audio as unconfirmed, so treat the audio parity as vendor-stated rather than independently checked.

• Inputs — Wan 3.0 accepts text, images, video, audio, and — new for the family — documents and web pages: PDF, PPT, DOCX, XLS, Markdown, TXT, up to 100 MB or 50 pages, one file or link per request. MiniMax H3 accepts text, image, video and audio references, up to 9 images, 3 video clips and 3 audio clips.

• Output control — MiniMax H3 offers first-frame, last-frame and first-and-last-frame control, plus reference-to-video. Wan 3.0 offers instruction-based editing, video extension, smart duration suggestion and adaptive aspect ratio. These are different feature lists, not different amounts of the same feature.

What the leaderboards say, and where they disagree

Artificial Analysis ranks both, and the pattern is not what the marketing implies.

• Text-to-video (with audio), 18 September 2026 — Wan 3.0 sits at #2 with an Elo of 1229 across 6,011 samples. MiniMax H3 sits at #4 with 1220 across 8,602. A nine-point gap on that board is real but modest.

• Image-to-video (with audio) — MiniMax H3 ranks #2 at 1181. Wan 3.0 ranks #6 at 1164. The open-weight model is ahead here.

• What that means — neither model dominates. Wan 3.0 is stronger from a text brief; MiniMax H3 is stronger from a still. The 30-second capability is the thing Wan 3.0 has that no leaderboard captures, because every board measures short clips.

A second source, Megaton's V-Benchmark v2.0 (updated 10 September 2026), scores Wan 3.0 at 81.6 overall and MiniMax H3 Max — the post-trained sibling of the base model, not the base itself — at 74.5. Different methodology, different answer, and the fact that the two benchmarks disagree is itself informative: these models are close enough that the ranking depends on what you measure.

Cost, at the resolutions you would actually buy

Prices here are per second of generated video, which is how both vendors bill, and both are quoted at list.

• Wan 3.0 — 0.3 yuan (about $0.05) per second at 480p, 0.6 yuan (about $0.10) at 720p, 1.2 yuan (about $0.20) at 1080p, per Alibaba. Artificial Analysis lists the 1080p configuration at $12.00 per minute. A 30% launch discount runs on Alibaba Cloud's Bailian and Qianwen AI platforms through 23 September 2026.

• MiniMax H3 — $0.08 per second at 768p and $0.13 per second at 2K on MiniMax's own platform, which Artificial Analysis renders as $7.80 per minute for the 1080p-equivalent configuration. Reference audio input is free; reference images are free for the first five.

• The draft trick does not save money — generating at 768p and regenerating to 2K costs $0.08 + $0.05 = $0.13 per second, exactly the 2K rate. It only pays off on takes you were going to discard anyway.

There is a real trap in comparing these two numbers directly: $0.08 and $0.20 are prices for different resolutions and different maximum lengths, so a raw per-second comparison flatters whichever model you were already leaning toward. The honest metric is cost per accepted clip, which includes the retries.

The routing angle, which is unusually concrete here

This is the rare matchup where the open-weight model is also the more available one.

MiniMax H3 is routable on OrcaRouter now, billed at the provider's list price with 0% markup — $0.08 per second at 768p, $0.13 at 2K, shown on the model page — with automatic failover, so a single serving path going down does not take the pipeline with it. Wan 3.0 is not on any routing platform: it is reachable through its own consoles — Model Studio (Bailian), its cloud console and the Wanxiang site — and through third-party creative tools such as Envato, Leonardo.Ai and ComfyUI. If your core requirement is "one key, many models, provider price," the H3 family is where that is already true.

That asymmetry cuts the other way too. Routing does not fix a licence. If you are in the US, UK, EU or South Korea and you intended to self-host MiniMax H3's weights, the route that actually works today is the hosted API — which is the same shape of access you get from Wan 3.0, minus the document input and minus the 30 seconds.

So which one

• You need more than 15 seconds, 1080p delivery, or to turn a document into a video — Wan 3.0. None of those has a MiniMax H3 equivalent.

• You need downloadable weights, first/last-frame control, or a 2K pipeline — MiniMax H3, provided your territory is not excluded; if it is, apply to MiniMax before you build on it.

• You want one API key and provider list pricing across a model family — MiniMax H3 is the half of this pair already routed, at list price with automatic failover.

• You are choosing by benchmark — stop. They trade places depending on whether you measure from text or from an image, and the gap on either board is smaller than the gap between your prompt and your neighbour's.

The thing to watch is whether Alibaba reverses course on weights. Every generation since Wan 2.2 has been more closed than the last, and Wan 2.2 remains the last Apache-licensed flagship — so the pattern points one way, but it is a pattern, not a promise. If a repository appears, this comparison gets rewritten.