MiniMax H3 Open Weights vs. ByteDance's Closed Seedance 2.5: Two Opposite Bets, Same Day

July 31 produced an unusual coincidence in Chinese AI video generation: MiniMax launched H3, and ByteDance launched Seedance 2.5, on the very same day. What makes it more than a coincidence is that the two companies picked opposite strategies. MiniMax announced it would release H3's model weights within days for developers to download, self-host, and fine-tune. ByteDance's Seedance 2.5, by contrast, said nothing about weights at all — same official-API-plus-consumer-subscription playbook it's always run. Same day, same category, two Chinese companies, two opposite calls — that's unusual enough that the South China Morning Post, Bloomberg, and Caixin all framed it as headline news within hours of each other, which is itself worth pausing on.

This piece won't stop at "one is open, one is closed." Behind that label sit two genuinely different business models: one trades free weights for developer mindshare and distribution, the other trades a fully closed stack for pricing power and product control. We'll walk through what H3 actually is and how it's built; how much of MiniMax's "open" claim has actually held up over the past year — and whether the terms have gotten looser or tighter; why ByteDance keeps choosing the opposite; how this fits alongside DeepSeek, Kimi K3, Qwen, and GLM once you zoom out to the whole Chinese landscape; and then the parts our readers care about most in practice — H3's real standing on Artificial Analysis's three independent leaderboards, how its pricing actually compares to peers, whether open weights bought it any legal cover, and whether mainland relay providers can serve it yet. Every figure below traces back to MiniMax's official blog, Artificial Analysis, Reuters, SCMP, TechTimes, BigGo Finance, and other cross-checked sources. Wherever sources contradict each other or can't be verified, we say so explicitly.

1. What H3 is: architecture and real specs

H3 is the latest generation of MiniMax's Hailuo line of video models. MiniMax's own blog frames it as an "open model that breaks the boundaries between tasks and modalities" — a single model that ingests text, images, video, and audio, and outputs a finished clip with the picture and soundtrack generated together, not stitched on afterward. That's the same broad technical bet ByteDance's Seedance line has made with its "unified multimodal joint audio-video generation" — both companies have converged on the same answer to "should sound and picture be generated separately," even though the implementations differ.

MiniMax's official blog names four architectural pieces: H3-VAE — a rewritten video tokenizer that MiniMax says delivers a "4x gain in effective sequence length," the technical basis for native 2K support; H3-Omni Transformer — folds understanding and generation workloads into a single pretraining framework instead of training separate expert models and merging them, which MiniMax says lifts end-to-end training throughput by nearly 30%; Contextual Omni Representation — uses language as a "generalizable bridge and interpreter" across modalities, compressing what would otherwise need roughly 100K tokens of reference material down to about 4K tokens of context; and In-Context Regeneration — generates high-resolution output directly from the original multimodal context, rather than upscaling a low-res draft with a separate super-resolution model. That's a reasonably complete technical description straight from the official blog — but one thing is conspicuously absent: MiniMax has not disclosed H3's parameter count or training data scale, unlike DeepSeek or Qwen, which routinely publish exact parameter counts. We flag that gap rather than guess a number.

On specs, multiple third-party sources cross-confirm: native 2K (2560×1440) at 24fps, plus a 768p tier still in invite-only closed beta; clip length of roughly 4-15 seconds (some sources say "5-15 seconds" — close enough in magnitude, differing slightly in exact wording); natively stereo audio generated in the same pass as the picture, not dubbed afterward; three entry points — text-to-video, image-to-video (first-frame generation with an optional last frame), and reference-to-video (mixed image/video/audio inputs); and editing capabilities including instruction-guided local edits, motion transfer between videos, and text/brand-element rendering. Reference-input limits vary slightly by source (some say up to 9 mixed references plus one audio track, others say up to 12 combined) — but either way it's in the single digits to low teens, far below the 50 reference inputs Seedance 2.5 launched with the same day. That's a real capability gap on the "long-form narrative, multi-asset fusion" front, and we note it plainly rather than talk around it.

2. "Open weights" — how much of that has MiniMax actually delivered this year?

The easiest thing to miss: as of this writing, H3's weights have not actually been released. Reuters' reporting is that MiniMax "plans" to release them "within days" — a pledge, not a completed fact — and MiniMax's own blog says it plans to open the weights "as soon as possible, subject to applicable laws and regulations." No H3 repository existed under MiniMax's Hugging Face organization as of this writing. Per the South China Morning Post, the license that will apply once weights do land is the MiniMax Community License: free for non-commercial use, and free for commercial use by organizations with under $20 million in annual revenue, with an attribution requirement. Structurally, that's close to Meta's Llama license family — not unconditional open source, but "open weights with a revenue gate."

And this isn't the first time MiniMax has used that kind of gated license — the gate itself has quietly tightened over the past year. Lay out the timeline: MiniMax M2 (its agentic/coding-focused MoE model), released October 2025, shipped under an unconditional MIT license and was called "the new king of open-source LLMs" by VentureBeat at the time. M2.5, in February 2026, kept the MIT license. But around April 2026, M2.7 quietly switched to a "Modified-MIT" license requiring written authorization for commercial use plus mandatory attribution — the community reacted immediately, with a Hugging Face discussion thread titled bluntly "Open source my ass — they are liars," and BigGo Finance characterizing the change as a "faux open-source" controversy. MiniMax's stated justification was preventing third parties from reselling degraded versions under its brand while preserving "commercial sustainability" — which, read plainly, means: now that it's a publicly listed company, the boundaries of free access need to be pulled back in for shareholders. By June 2026's flagship text model M3 (428B total / 23B active parameters MoE, 1M context — also tracked on this site's own model leaderboard as "the first domestic open-weight model to combine multimodal and long-context support"), the weights did land on Hugging Face, but under that same gated "minimax-community" license, and MiniMax still hasn't published training code or inference operators — which is why Open Source For You's headline on M3 read "stops short of full open source commitment."

3. Why ByteDance stays closed: what 200 million daily users buys you

ByteDance's logic here is straightforward once it's laid out plainly, even though it rarely gets stated in so many words. Multiple outlets have summarized ByteDance's four AI priorities for 2026: world models (targeting at least one release by year-end benchmarked against Google's Genie 3, reportedly about 10% behind state of the art internally, with 3-4x the data budget of peer companies); video generation (the Seedance line, addressing what's described internally as an "anti-scaling law" problem); coding (ByteDance has moved from voluntary to mandatory internal use of its own Seed models across multiple business units, explicitly benchmarking against Claude Code's roughly $2.5B annualized revenue); and monetizing Doubao — which already has 200+ million daily active users and is shifting from free to paid tiers running up to 500 RMB a month, with overseas sibling app Dola targeting 30 million daily users by year-end, up from 10 million at the end of 2025.

Put those four together and the pattern is a classic vertically integrated playbook: ByteDance doesn't need open weights to buy developer mindshare, because it already owns the distribution — 200+ million daily users across Doubao and Jimeng, plus an enterprise cloud arm (Volcano Engine/BytePlus). Model value gets captured directly inside its own consumer products, and internally mandating use across ByteDance's own business units is itself a form of forced distribution. That's the same logic OpenAI, Google, and Anthropic run — subscriptions plus a closed enterprise API — just executed by a Chinese consumer-internet giant instead. In fairness, ByteDance isn't closed across the board: its Seed team has published genuinely open-weight research and tooling models on Hugging Face, including Seed-OSS (Apache 2.0, trained on roughly 12 trillion tokens) and UI-TARS (a GUI-agent vision-language model, released in multiple parameter sizes). But the products that actually carry ByteDance's commercial weight — the Doubao Pro line, the Seedance line — have been closed-API-only from day one. That distinction matters and shouldn't be flattened into "ByteDance never open-sources anything."

4. Zooming out: China's open and closed camps

Looking only at MiniMax and ByteDance, it's tempting to read this as just another instance of the tired "China open-source vs. Western closed-source" framing. It's actually the opposite: this is two Chinese companies choosing opposite paths, which is exactly why SCMP, Bloomberg, and Caixin all zeroed in on the contrast within hours of each other — the open-vs-closed dividing line is shifting from "China vs. Silicon Valley" to a fight happening inside China's own AI industry.

Zoom out further and the landscape gets more interesting. On the open-weight side, alongside MiniMax sit companies we've reviewed on this site before: DeepSeek (V3/R1/V4 all shipped under MIT, with unusually high transparency about training details for the category), Moonshot's Kimi K3 (2.8 trillion parameters, described by Bloomberg as one of the largest open-weight AI systems in the world), Alibaba's Qwen (Apache 2.0 across the full size range from 0.5B up), and Zhipu's GLM (GLM-5.2 also under MIT) — the group that's carried the "China wins through open-source distribution" narrative for the past two years. That camp isn't monolithic either: per ciw.news, Zhipu's own recent GLM-5 Turbo release went proprietary, showing that even a long-time open-source standard-bearer can pull back under commercial pressure. And within this camp, MiniMax currently sits on the more conservative end of licensing terms — a revenue gate, an attribution requirement, and undisclosed training code put it further from "fully open" than DeepSeek's or Qwen's cleaner open-weight releases.

On the closed side, alongside ByteDance sit the usual Western labs covered elsewhere on this site — OpenAI, Google, Anthropic — whose business logic mirrors ByteDance's almost exactly: subscription consumer products plus a closed enterprise API, where the model itself is the core asset not to be given away for ecosystem points. Put the whole map together and a sharper pattern emerges: the open-vs-closed choice mostly tracks whether a company already owns a large enough consumer distribution channel. Own one — ByteDance's 200 million daily Doubao/Jimeng users — and closed monetization makes sense. Don't have one yet, or you're still fighting for developer mindshare — DeepSeek, Kimi, MiniMax in its earlier days — and open weights buy you an ecosystem. MiniMax opening H3's weights is, in that light, an extension of the exact playbook that already worked for its text model (M2) into video generation, a category that's been almost uniformly closed-source until now. Reuters quoted MiniMax's own framing directly: "closed-source models have long dominated video generation, with slower iteration and a less open ecosystem than fields like large language models."

5. Performance review: H3's real rankings on Artificial Analysis's three leaderboards

Strategy aside, the question our readers care about most in practice: is H3 actually good? Rather than repeat MiniMax's own marketing claims, here's what independent benchmarking firm Artificial Analysis has published as of this writing — its methodology has real users blind-vote between outputs from different models, converted into Elo scores, and it's one of the more widely cited third-party benchmarks in video generation right now.

Leaderboard	H3's rank / score	Ahead of / behind
Text-to-Video	#2, Elo 1242	Behind Gemini Omni Flash (#1, 1245); ahead of Seedance 2.0 (#3, 1225), Alibaba's Wan2.7 (#4, 1163), and all Kling 3.0 tiers (ranked #7-13, 1092-1113)
Image-to-Video	#3, Elo 1185	Behind Seedance 2.0 and Gemini Omni Flash (we could not verify a consistent source for which of those two ranks higher, so we don't assert an order between them)
Video Editing (with audio)	#1, Elo 1130 (from 5,043 blind-preference samples)	Artificial Analysis's top-ranked video-editing model, period — H3's one unambiguous first place

Taken together, H3's position is fairly clear: not a clean sweep, but top-three across all three dimensions, and outright first in video editing. The text-to-video leaderboard is particularly notable — H3 (1242) edges out Seedance 2.0 (1225, and note this is a comparison against 2.0, not the same-day-launched 2.5, since Artificial Analysis has no independent entry for Seedance 2.5 as of this writing — a gap our own Seedance 2.5 review flagged too), trailing the leader Gemini Omni Flash by just 3 points. That's a genuinely tight gap, putting H3 in the top tier on the most basic video-generation capability. Image-to-video is where it's weaker — ranked third, behind both Seedance 2.0 and Gemini Omni Flash — meaning "extend a still image into video" isn't where H3 currently leads.

6. Pricing: is H3 really "less than a third" of mainstream rivals?

MiniMax's own blog claims H3's per-second price at 2K is "less than a third of mainstream models," and its 768p tier is "less than half the price of mainstream models' 720p." Here are the actual verifiable numbers before judging whether that claim holds up.

Model	Resolution tier	Known price	Source
MiniMax H3	2K	$0.13–0.14/second	Official pricing, OpenRouter
MiniMax H3	768p (invite-only beta)	$0.09–0.10/second	Official pricing (sales contact required)
Seedance 2.0	720p	~$0.22–0.24/second	Third-party review site
Seedance 2.0	1080p	~$0.374/second ($22.45/min)	Third-party channel (Atlas Cloud)
Kling 3.0	1080p	~$0.336/second ($20.16/min)	Third-party channel (Atlas Cloud)
HappyHorse-1.1	—	~$0.165/second ($9.90/min)	Third-party channel (Atlas Cloud)

By this table, H3's 2K rate of $0.13-0.14/second against Seedance 2.0's 1080p rate of $0.374/second is close to a two-thirds discount; against Kling 3.0's $0.336/second, similar magnitude. "Less than a third of mainstream rivals" holds up reasonably well against these two specific reference points. But one thing needs to be stated plainly: the comparison here is against Seedance 2.0, not Seedance 2.5 — the model that actually launched the same day as H3. Our own Seedance 2.5 review already covered this in detail: ByteDance has not published official developer-facing API pricing for Seedance 2.5 at all, and third-party estimates for it contradict each other and don't even agree on the order of magnitude. In other words, "H3 is cheaper than Seedance" can currently only be verified as "H3 is cheaper than Seedance 2.0, the previous generation" — H3 and its actual same-day rival Seedance 2.5 cannot be price-compared at all right now. That's a detail easy to lose in a press-release summary, and we think it's worth spelling out explicitly.

7. Copyright: open weights bought no legal cover

There's an intuitive assumption that open models are "more transparent" and therefore carry less copyright risk. The real situation behind H3 argues the opposite. Multiple outlets report that Disney, Universal Pictures, and Warner Bros. Discovery jointly filed a copyright infringement lawsuit against MiniMax on September 16, 2025, in the US District Court for the Central District of California, alleging that its Hailuo AI video service (H3's predecessor product) trained on unauthorized copies of their copyrighted characters — the complaint reportedly cites recognizable outputs resembling Spider-Man and Darth Vader. More significantly, per AI Weekly and other coverage, a federal judge has already denied MiniMax's motion to dismiss the case, ruling the three studios have plausible claims — the case is now in discovery, covering training-data documentation, model architecture decisions, and Hailuo's US revenue figures.

Set that against what our own Seedance 2.5 review found about ByteDance's situation: ByteDance has so far only received a cease-and-desist letter from the MPA, and no studio has actually filed suit in US court yet (partly attributed to the time cross-border service of process takes). MiniMax's case, by contrast, isn't just filed — its motion to dismiss has already failed, putting it materially further along in real litigation exposure. In other words, on the copyright front, the open-weight company (MiniMax) is currently deeper into actual court exposure than the closed-source one (ByteDance). Worth flagging for any team planning to use H3 for commercial content, especially anything that might resemble existing film or character IP: whether a model is open source and how much copyright risk it carries are two completely independent questions — being open doesn't mean being safer.

8. Mainland relay access: providers moved faster than the official weights did

The question that matters most practically for our readers: can mainland relay providers serve H3 yet? Unlike Seedance 2.5, where the official API itself wasn't open and providers were collectively waiting, H3's situation is reversed — even though MiniMax's promised weight release hasn't landed as of this writing, H3's inference API itself is already live. OpenRouter already lists an entry for "MiniMax: H3" (model ID hailuo-3) priced at $0.13/second, with a listed availability date around July 29 — two days ahead of the July 31 official blog announcement, likely an early-access rollout to select channels. Beyond OpenRouter, several overseas video-inference platforms including fal.ai, WaveSpeed AI, ModelsLab, and EvoLink already have dedicated H3 product pages with clear pricing.

Among providers already tracked on this site, Atlas Cloud (the video-model-focused aggregator we noted in our Seedance 2.5 review, covering Seedance 2.0, Kling V3.0, Alibaba's Wan 2.7, Vidu, and Grok Imagine) has already launched a dedicated MiniMax H3 product page with published pricing. That means even though H3's promised open weights haven't actually materialized, it has reached usable status through relay/inference platforms faster than its same-day rival Seedance 2.5, whose official API wasn't even open at the time of our review. Separately, FlowBar — another provider already in our database — lists "minimax" among its covered model families, but this appears to reference MiniMax's text-model (M-series) lineup rather than confirming coverage of the H3 video model specifically, so verify with the provider directly before assuming it's covered.

Practical advice for anyone integrating: don't assume "supports MiniMax" means "supports H3" — the same principle we've stressed before applies here. Confirm the exact model ID (hailuo-3 or minimax-h3), confirm whether you're getting the 2K or 768p tier (a 30-40% price gap between them), and confirm whether the provider is directly forwarding to MiniMax's official API or plans to self-host once the real open weights land — the latter could mean meaningfully different latency and reliability depending on each provider's own hosting setup. Keep an eye on our AI API relay provider comparison as providers update their model coverage.

9. Who should care now, who should wait

• E-commerce/advertising/gaming teams needing low-cost, fast-iterating video content → H3's 2K pricing has a clear edge over Seedance 2.0 and Kling 3.0, and it's already reachable through OpenRouter, Atlas Cloud, and similar channels — worth testing now.
• Creators needing long-form, multi-asset-fusion content → H3's reference-input ceiling is far smaller than same-day-launched Seedance 2.5's 50 inputs; if your core need is multi-minute narrative continuity, Seedance 2.5's product direction fits better, though its official API pricing remains an open question.
• Teams planning to self-host or fine-tune the open weights → the weights haven't actually shipped yet — "within days" is a pledge, not a fact — and even once they do, organizations above $20M in annual revenue need a separate commercial authorization, not unconditional free commercial use. Confirm the license terms before planning a deployment timeline around it.
• Teams planning commercial content that might resemble existing film/character IP → MiniMax (Hailuo) is already deep into a real federal lawsuit with its motion to dismiss denied; copyright risk assessment before commercial use matters more than ever, and "it's open source" is not a reason to relax that scrutiny.
• Mainland developers → H3's inference API is already reachable through several overseas aggregators and at least one provider tracked on this site, but confirm the exact model ID and resolution tier — don't assume "supports MiniMax" equals "supports H3."

10. Conclusion

That MiniMax H3 and ByteDance's Seedance 2.5 launched the same day on opposite strategies is, in itself, more worth remembering than either model's individual benchmark score — it shows that "open or closed" is far from a settled question even within Chinese AI, and two companies with very different scale and distribution reached opposite conclusions from their own respective business logic: ByteDance already has 200 million daily users across Doubao and Jimeng and a proven closed-monetization path, so it has no need to trade weights for ecosystem; MiniMax is still fighting for developer mindshare in a video-generation category that's been almost uniformly closed, and open weights is its most cost-effective way to buy distribution.

But this shouldn't be simplified into "open deserves more support, closed is more conservative." H3 itself turned in a solid scorecard — two top-three finishes and one outright first place across Artificial Analysis's three leaderboards. But its "open weights" remain, as of this writing, an unfulfilled pledge, its license carries a revenue gate, and MiniMax's copyright litigation is materially further along than ByteDance's — all facts that belong in a complete picture of this standoff, not footnotes to skip past.