TLDR
- Wan 3.0 (public beta 6 August 2026; Alibaba Cloud called it officially released 24 August 2026): Alibaba's all-in-one video model. Public story: up to 30 seconds in one pass, text / image / audio / video / documents as input, 480P / 720P / 1080P, native sound with the picture (Model Studio blog; Alizila; @alibaba_cloud).
- MiniMax H3 Max (hosted speed window from late August 2026): a hosted, post-trained variant of open MiniMax H3 — not MiniMax's official product name. Public story: 5–15 second clips, 480P / 768P on the hosted surface, native audio, built to return a take faster than you watch it (Artificial Analysis; Design Arena; H3 Max).
- Decision rule: need a single 10–30s scene with dialogue in the same pass, 720P/1080P, or a document-led brief on the official family → Wan 3.0 on this site. Need many 5–15s drafts in one sitting at 480P/768P → try H3 Max.
- On this site: the generator is Wan 3.0. There is no H3 Max button here.
Key Takeaways
- These are not the same MiniMax. Open MiniMax H3 is MiniMax's 31 July 2026 omni-modal model: up to 15 seconds, 2K path, native stereo, weights public (MiniMax H3; open-source note; Reuters). H3 Max is the later hosted speed variant of that base. MiniMax's own blog still names H3, not Max.
- Clip-length public claims differ by a full beat. Wan 3.0's launch materials center on 30 seconds in one generation, double Wan 2.7's 15-second cap (Model Studio; TechNode). H3 Max sits on the H3 length band: 5–15 seconds (Artificial Analysis).
- Resolution ceilings are not the same knob. Wan 3.0 documents 480P / 720P / 1080P — no 4K in the launch spec (Model Studio). Hosted H3 Max public cards stop at 768P; open H3 still owns the 2K regenerate path (H3 open-source).
- Independent boards are a photo finish, not a crown. On 2 September 2026, Artificial Analysis' text-to-video (with audio) board read: Gemini Omni Flash 1238, Wan 3.0 1237, MiniMax H3 Max 1235 — overlapping confidence bands, not a locked #1 (T2V board). The same day, Wan 3.0 led video editing (with audio) at Elo 1189 (editing board).
- Native audio is table stakes on both official stories. Alibaba describes multilingual voice with the picture (Alizila; GitHub Wan 3.0). MiniMax H3 ships 32 kHz stereo (open-source note). Listen every take. Alibaba also flags audio texture as still improving (Model Studio).
- Same-prompt evaluation on one hosted surface still beats a social montage. Run the prompts below on this site first. For the H3 Max half of the same brief, open h3max.ai.
Head-to-head: public axes (not a ranking)
Read this table in 30 seconds. Cells are public positioning / documented features, not lab scores we ran.
| Axis (your job) | Wan 3.0 (Alibaba) | MiniMax H3 Max (hosted H3 variant) |
|---|---|---|
| Primary public pitch | 30-second scenes from almost any input, including documents (Model Studio) | Faster hosted take on MiniMax H3: prompt follow + clips that return sooner than playback (Design Arena) |
| What it is | Alibaba's Wan-family all-in-one video model (GitHub) | Hosted post-trained MiniMax H3 — not MiniMax's open H3 drop, not a MiniMax-named official SKU (AA) |
| Public window | Public beta 6 August 2026; official-release post 24 August 2026 (Alizila; @alibaba_cloud) | Arena/speed posts late August 2026 (Design Arena) |
| Single-pass duration | Up to 30 seconds, with smart duration and extension (Model Studio) | 5–15 seconds (AA) |
| Resolution narrative | 480P / 720P / 1080P (Model Studio) | Up to 768P on public H3 Max cards; 480P for cheap drafts. Open H3 still has 2K (H3 open-source) |
| Speed narrative | No official end-to-end second count in the launch post. Draft at 480P, finish at 1080P is the stated workflow (Model Studio) | Design Arena: MiniMax H3 quality at more than 50× the speed of the H3 base (Design Arena). August 2026 hosted benchmarks: a 5-second 768P clip in under 3 seconds |
| Inputs / control | Text, image, audio, video, documents and web pages; edit visuals, plot, and dialogue (Model Studio; Alizila) | Text-to-video and start-frame image-to-video on the hosted Max surface. Omni reference/edit stays with open H3, not Max (H3 post) |
| Native audio | Voice with the picture; Alibaba notes audio texture is still maturing (Alizila; Model Studio) | Native audio on 5–15s clips; open H3 specifies 32 kHz stereo (H3 open-source) |
| Open weights | No. Open Wan weights stop at Wan 2.2. Wan 3.0 is a hosted API model (GitHub) | Max weights not public in sources we have. Open H3 weights are a different download (H3 open-source) |
| Independent board (read 2 Sep 2026) | T2V with audio Elo 1237 (band 1–4); video editing Elo 1189, #1 that day (T2V; editing) | T2V with audio Elo 1235 (band 1–4), listed as a post-trained MiniMax H3 build (T2V) |
| Try path | Homepage generator on this site | Not hosted here. Generate on h3max.ai |
How to use the table: if the brief is “one 20–30s scene, voices included, 1080P keeper,” weight Wan 3.0 on this site. If it is “twelve 8-second variants before lunch, 768P is enough,” weight H3 Max speed.
Criteria, not crowns · editorial cover for this comparison · generated for the Wan 3.0 AI blog
Family / timeline (keep the lanes straight)
| Date | Event | Source |
|---|---|---|
| 2026-07-31 | MiniMax H3 (open omni-modal video) launches | MiniMax H3; Reuters |
| 2026-08-03 | MiniMax H3 weights / system card public | H3 open-source |
| 2026-08-06 | Wan 3.0 public beta: 30s, documents, 1080P | Model Studio; Alizila |
| 2026-08-10 | Independent write-up of the Wan 3.0 beta | TechNode |
| 2026-08-24 | Alibaba Cloud posts Wan 3.0 as officially released | @alibaba_cloud |
| 2026-08-26 | H3 Max arena/speed window: Design Arena | Design Arena |
| 2026-09-02 | AA T2V-with-audio: Gemini 1238 / Wan 1237 / H3 Max 1235 | T2V board |
Wan 3.0 extends Alibaba's Wan line (Wan 2.7 capped at 15 seconds) into a longer, all-in-one generation (Alizila; official site). On this site, start at the homepage generator.
MiniMax H3 Max sits on top of MiniMax H3. Do not treat Max as “H3 but 2K,” and do not treat open H3 as “the fast hosted one.” If you want the fast hosted variant, the browser path is h3max.ai — not this generator.
What the official demos actually look like
Launch and hosted samples — not a timed A/B on this site.
Wan 3.0 official demo · a long single take (cherub on a spiral stair) · hosted on this site · Alibaba GitHub
MiniMax H3 Max sample · a short hosted take · on h3max.ai · not a lab A/B vs Wan 3.0
Scenario winners (if / then)
1. Twelve UGC drafts before lunch
If you need a pile of 9:16 or 16:9 5–15s takes the same afternoon, then weight the H3 Max speed story (Design Arena).
Try path: generate those drafts on H3 Max. Use this site when you want a longer Wan-family preview of the same idea instead.
2. One-take ad, ~15–30 seconds, voices included
If the brief is a continuous camera move that has to land as one file, with dialogue already in the pass, then Wan 3.0's public 30-second + native-sound story is the axis to test first (Model Studio; Alizila).
On this site: write the whole beat in one paragraph on the homepage generator. Default length is 5 seconds — raise it to match the action.
3. Locked product still → motion
If packaging is already approved as a still, then start image-to-video on both stacks. Wan 3.0 documents first-frame (and first-and-last-frame) image-to-video (GitHub). H3 Max's hosted surface is built around a start frame plus an optional end frame (h3max.ai).
Split try path: freeze the still, then animate here; run the same still on H3 Max if you need more motion variants per hour.
4. Deck, PDF, or webpage → narrated film
If the source is a document or URL, then that is a Wan 3.0 family claim: Model Studio lists doc / xls / ppt / pdf / txt and web pages, one file or link, ≤100MB / 50 pages (Model Studio).
On this generator today: start from text, a first frame, and image / video / audio references. Do not assume a PDF upload button is on this page. H3 Max does not advertise document-to-video.
5. Open weights, 2K, or omni references
If you need downloadable H3, 2K regenerate, or mixed image/video/audio references on MiniMax's own terms, then that is MiniMax H3, not H3 Max (H3 open-source). Max is the hosted speed lane.
On this site: you still evaluate the language of the brief in the homepage generator so a soft prompt does not get blamed on either model.
What We Know vs. What We Don't
| We know (sourced) | We don't know / won't claim |
|---|---|
| Wan 3.0 public beta 2026-08-06; Alibaba Cloud official-release post 2026-08-24 (Model Studio; @alibaba_cloud) | A permanent global #1 between Wan 3.0 and H3 Max across every job |
| MiniMax H3 launch 2026-07-31; H3 Max public arena window late August 2026 (Reuters; Design Arena) | That MiniMax itself shipped a product named “H3 Max” — official pages through this writing still talk H3 |
| AA T2V-with-audio on 2 Sep 2026: Wan 1237, H3 Max 1235, Gemini 1238 (T2V) | That Elo will hold next week, or that it predicts your keeper rate |
| Wan 3.0 led AA video editing (with audio) at Elo 1189 that day (editing) | A timed A/B we ran on wan30ai.org vs h3max.ai |
| This site hosts Wan 3.0, not H3 Max | Cross-channel $/second shopping tables (not compared here) |
| This generator: text, first/last frame, image/video/audio references; 2–30s; 480P / 720P / 1080P (default 480P) | That this page currently accepts a PDF or webpage as the only input |
How to evaluate on this site (same-prompt protocol)
- Open the Wan 3.0 generator.
- Run Prompt A, B, and C below unchanged (change only product names you truly need).
- Score each take (0–2 each): cut clarity, subject identity, motion realism, audio fit, crop survival.
- Keep any clip ≥7/10. Rewrite only the failing shot line.
- For the H3 Max half of the same test, paste the same three prompts on h3max.ai — change only the host. Expect H3 Max to cap around 15 seconds; shorten A if a take gets truncated.
- Optional: lock a still, then image-to-video both hosts.
Credit cost on this site follows resolution × seconds (floor applies on the shortest 480P run). Check pricing before a long 1080P keeper. This article does not compare anyone else's per-second menu.
Prompt A — multi-shot social UGC
Vertical 9:16 UGC ad, about 10 seconds, daylight kitchen.
Shot 1 wide: young adult in a blue hoodie pours iced coffee into a clear glass, casual phone-tripod energy.
Shot 2 medium: glass fills, ice clinks, condensation visible, same hoodie and kitchen tiles.
Shot 3 close-up: satisfied sip, natural smile, soft window light, no logos.
Natural kitchen ambience and ice clink SFX, no music bed, no captions, no watermark.
Prompt B — dialogue explainer beat
Horizontal 16:9 product explainer, about 12 seconds, clean home office.
Medium shot: friendly presenter in a charcoal sweater speaks to camera,
holds a small white wireless earbud case, clear enunciation, natural hand gestures.
Background: blurred bookshelf, soft daylight.
Native dialogue: a short line about "all-day battery," ambient room tone only, no music, no on-screen text, no watermark.
Prompt C — one-take street beat
Horizontal 16:9 cinematic one-take, about 20 seconds, golden-hour street.
Handheld gimbal push: a courier in a red jacket exits a cafe doorway, weaves past two cyclists,
then stops at a scooter, checks a phone, and rides off down a tree-lined avenue.
Continuous camera; same jacket and bag throughout; natural city ambience, scooter ignition, light traffic, no music, no logos, no watermark.
Prompt C is the Wan 3.0-shaped brief (length past 15 seconds). On H3 Max, cut it to 15 seconds or treat a truncated ending as a known cap, not a model failure.
Decision rule (one sentence)
Choose Wan 3.0 on this site when you want a browser 2–30s scene + native audio in the Wan family, and you will judge keepers yourself on the homepage generator.
Choose MiniMax H3 Max when the bottleneck is iteration speed on 5–15s, 480P/768P clips — open h3max.ai. Do not expect Max to be open H3, and do not expect this site to generate Max.
FAQ
Is MiniMax H3 Max the same as MiniMax H3?
No. H3 is MiniMax's open omni-modal model (4–15s, 2K path, public weights) (H3; open-source; Reuters). H3 Max is a later hosted, post-trained variant aimed at speed and prompt follow, discussed from late August 2026 (AA; Design Arena). MiniMax's own channels still describe H3, not Max, in the pages we checked.
Is H3 Max “better” than Wan 3.0?
Not as a universal rank. Public materials show different jobs: fast 5–15s hosted drafts vs longer one-pass scenes with a higher resolution ceiling. Better = fits the brief after the same three prompts.
How long can each generate in one pass?
Wan 3.0 launch coverage: up to 30 seconds, with smart duration and extension (Model Studio). H3 Max public cards: 5–15 seconds (AA). Host knobs can still cap what you see. This generator defaults to 5 seconds.
Did Wan 3.0 take #1 on Artificial Analysis?
On 2 September 2026 it did not lead text-to-video with audio. Gemini Omni Flash sat at 1238, Wan 3.0 at 1237, H3 Max at 1235 — overlapping bands (T2V). Wan 3.0 did lead video editing with audio that day at 1189 (editing). Rankings move. Date the number.
Where do I actually generate MiniMax H3 Max?
On h3max.ai — start at the generator. This site does not host that model.
Can I try Wan 3.0 without installing local tools?
Yes. Open the homepage generator in the browser. Plans and credit packs are on the pricing page.
Do both models generate audio?
Yes in their public stories (Alizila; H3). Native audio ≠ perfect lip sync. Alibaba says audio texture is still being refined (Model Studio). Headphones, every take.
Can Wan 3.0 output 4K?
No in the launch spec. Output is 480P, 720P, or 1080P (Model Studio). H3 Max hosted cards stop at 768P; open H3 reaches 2K through regenerate (H3 open-source).
Is Wan 3.0 open source?
No. Alibaba has not released 3.0 weights; the downloadable open Wan line stops at Wan 2.2 (GitHub). MiniMax did open H3-Base under a community license — that is a different model (H3 open-source).
Should I compare prices across platforms here?
No. This article skips $/second shopping tables. Decide on job fit and keeper rate, then price on the surface you actually ship from. This site's credit math is on the pricing page.
About Nia Hart
Nia Hart is an AI video model analyst who tracks Alibaba Wan releases and turns public model facts into clear try paths for creators on Wan 3.0 AI.