MiniMax H3: video generation at 0.8 yuan a second

The video generation chart got a new leader on July 31, and it's priced like a SaaS product, not a frontier model: MiniMax H3 generates up to 15 seconds of 2K video with native dual-channel audio at 0.8 yuan per second. Roughly a third of what comparable flagships charge.

H3 is a full-modal generation model, which is a mouthful until you see it work: text, image, video, and audio in any combination, one video out with synchronized sound. MiniMax calls it a "kill line" for the category. We usually ignore that kind of marketing, but the price math and the leaderboard placements deserve a compiled look.

What shipped

H3 accepts any combination of text, image, video, and audio as input, and outputs video with native dual-channel audio, up to 15 seconds at 2K [1][2]. Two capabilities stand out in the release materials: V2V Motion Transfer, which transfers motion from one video to another subject, and a captioning pipeline that compresses most 100K-token inputs to about 4K tokens for inference [2][3]. The latter is how the cost stays low: a high-compression tokenizer plus heterogeneous training, load balancing, and GPU utilization tuning [2]. Dual-channel audio matters more than it sounds: the output carries synchronized voice and music in one pass, which is what turns a model call into a usable ad spot or product clip instead of a silent render you have to score separately.

Target use cases per MiniMax: advertising, e-commerce, product design, UI/UX, and games [3]. This is a commercial playbook, not a research demo.

Where it ranks

• Artificial Analysis video editing eval: first globally [4].
• LMArena image-to-video leaderboard: first [4].
• Hugging Face trending: #1, above DeepSeek V4 Flash [4]. H3 also picked up likes from Stable Diffusion's founder Emad Mostaque and a16z partner Justine Moore [4].

The HF trending placement is the one I'd weight. Video benchmarks are young and thinly populated; trending is a real-time signal of what developers actually pulled down.

The 15-second ceiling is the honest constraint. That's long enough for ad cuts, product shots, and UI motion previews, the exact use cases MiniMax lists. It's short for storytelling. If your pipeline needs 30-second scenes with continuity across shots, H3 is not there yet, and no benchmark on this board measures continuity anyway.

The price math

0.8 yuan per second at 2K, about a third of flagship rivals [2]. At 15 seconds that's 12 yuan, roughly $1.70, for a full dual-channel clip. For comparison, that price level is what made text LLM APIs boring, and it's arriving in video, where the incumbents still bill per-clip like it's 2023.

The stock market noticed before we did: MiniMax shares rose over 14% the day after the release, and the company was added to the Stock Connect (HK) list days later [2][5].

The open-weights promise, and the caveat

Update (Aug 7): The weights landed. MiniMaxAI/MiniMax-H3 went up on Hugging Face around August 3 (12,102 downloads, 2,675 likes at our API capture), and the license is the catch: the H3 Community License Agreement excludes the EU, UK, South Korea, and the US. Our News Desk follow-up covers the license in full.

MiniMax said on July 31 that weights would be open "within days," enabling local deployment for enterprises [1][2]. At our August 6 capture we hadn't found them on Hugging Face, and we flagged that as the main open item — see the update above for what happened next. That line, between "we read about a model" and "we can run it," is now closed: the weights are up, with a territory-gated license.

The other caveat is vaguer. A CSDN technical review notes the model has "certain shortcomings" without naming them [6]. We do not run with claims we cannot source, so that line stays in the archive as an unverified flag. If a concrete gap appears, we'll update the row.

Our take

H3's real move is the price, and the price is a strategy, not a cost accident. MiniMax engineered the tokenizer and the serving stack to make video cheap, then published the numbers. The market's reaction, HF trending #1 and a 14% stock pop, says the category was waiting for exactly that.

Here's what gets me: video generation just crossed the same line that text did in 2023. When a generation task gets cheap enough, it stops being a demo and becomes a pipeline component. Ads, product shots, UI motion previews, those are workflows people will now build on top of H3's API. The open-weights promise, when it lands, makes it a self-hostable pipeline component, which is how you win developers, not just API users.

The pressure this puts on the rest of the market is the part nobody's saying out loud. Every video model priced above H3 now has to justify the difference on quality or features, and quality claims are exactly what independent evals are slowest to settle.

Risks and what we're watching

• The weights. Resolved — see the update above: H3-Base is on Hugging Face (Aug 3), license excludes the EU, UK, South Korea, and the US. The open question moved from "will they ship" to "who's licensed to run them."
• The unspecified shortcomings in community reviews. Vague criticism is still a signal. We want specifics.
• Benchmark thinness. AA's video editing eval and the Arena image-to-video board are new; placements can reshuffle fast.
• Model quality at the price point. Cheap video that needs heavy editing is a different product than cheap video that ships.

Bottom line

MiniMax H3 leads the video charts on Artificial Analysis and Arena, topped Hugging Face trending, and priced 2K video with audio at a third of the flagship rate. It's the strongest video-generation value we've tracked. The open-weights promise is settled — H3-Base shipped on August 3 with a license that excludes the EU, UK, South Korea, and the US, which is a different kind of open. We'll keep the leaderboard row current as deployment stories land.

Same number on our leaderboard → MiniMax H3 row

Sources

1. IT Home (H3 release, full-modal, 15s 2K, captioning), ithome.com/0/983/957.htm (published 2026-07-31, captured 2026-08-06, media report)
2. Wall Street CN via 163 (pricing 0.8 yuan/sec, cost engineering, stock +14%), 163.com/dy/article/L3623H0E05198NMR.html (published 2026-07-31, captured 2026-08-06, media report)
3. Leiphone (kill line, 1/3 flagship price), leiphone.com/ (published 2026-07-31, captured 2026-08-06, media report)
4. Leiphone (AA video editing #1, Arena image-to-video #1, HF trending #1, Mostaque/Justine Moore likes), leiphone.com/category/industrynews/aiUMBoeUYbi8fX4x.html (published 2026-08-05, captured 2026-08-06, third-party eval + media)
5. 36Kr newsflash (Stock Connect inclusion), 36kr.com/newsflashes/3926617883212168 (published 2026-08-05, captured 2026-08-06, media report)
6. CSDN (technical review, "certain shortcomings", weights promise), blog.csdn.net/techforward/article/details/163372415 (published 2026-07-31, captured 2026-08-06, community)

Disclosure: no sponsors, no affiliate links. Compiled from public sources with capture dates; we did not run H3 ourselves. Weights confirmed on Hugging Face via API capture; license text quoted as captured.