MiniMax H3 Launches: 2K AI Video With Native Audio

MiniMax H3 is a video-generation model — let’s clear that up before anything else. Released July 31, 2026, H3 is the successor to MiniMax’s Hailuo video line, not a new version of the M3 language model the company open-sourced in June, and “Hailuo 3.0” is a community nickname that MiniMax itself never uses. The model ships with 2K output, natively synced audio, and reference-driven generation from up to twelve input files.

The launch matters for two reasons. First, the spec sheet is genuinely aggressive: 4–15 second clips at 2K with dialogue, ambience, and music generated in the same pass as the picture, at $0.13 per second pay-as-you-go — a price MiniMax claims comes in under a third of mainstream rivals’ cost for 2K output. Second, it landed the same day as ByteDance’s Seedance 2.5, turning July 31 into a head-to-head skirmish in China’s AI-video price war.

This guide untangles the naming confusion, walks the full spec and reference-input system, computes what a finished clip actually costs (MiniMax only publishes unit rates), and separates what shipped from what was merely promised — starting with the weights, which as of publication exist as a pledge, not a download.

Launch-day coverage has been conflating three names, so here is the clean split. H3 is the new video model — the general-purpose multimodal video generator MiniMax released on July 31, 2026, exposed in the API under the literal model ID MiniMax-H3. M3 is a different product entirely: the open-weight, 1M-context language model MiniMax released on June 1, 2026, which we covered in our M3 release analysis and benchmarked in our M3 vs Opus vs GPT-5.5 agentic-coding comparison. And “Hailuo 3.0” is a community label — H3 does succeed the Hailuo product line, but MiniMax’s own docs and its official X account use only “H3” and “MiniMax H3.”

The distinction is practical, not pedantic. Teams searching for H3 benchmarks will find M3 coding scores; procurement teams pricing “the new MiniMax model” may quote the wrong product line entirely. When a vendor ships two flagship releases two months apart with near-identical names, the burden of disambiguation lands on everyone downstream — treat any “Hailuo 3.0” claim you encounter as secondhand until you trace it to MiniMax’s own H3 documentation.

H3 outputs clips of 4 to 15 seconds (integer values only) at 2K resolution — the only tier that is generally available at launch. Audio is generated natively: dialogue, ambience, and music are produced in the same generation pass as the picture, which Reuters describes as 2K video “with native stereo sound.” Multiple independent third-party write-ups report 24fps output; MiniMax’s own docs specify resolution, duration, and aspect ratio but omit an explicit frame-rate figure, so treat 24fps as reported rather than vendor-confirmed.

Prompts are required and capped at 7,000 characters per request. Aspect ratio is adaptive by default for image-to-video, while text-to-video requires an explicit ratio from a fixed set — 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16 — with the width-to-height ratio constrained to fall between 2:5 and 5:2. Beyond straight generation, Reuters reports that H3 can edit existing content and transfer movement from one video to another using instructions and reference material — capabilities that push it toward production-editing territory, not just clip synthesis.

The reference-generation system is the most distinctive part of the spec sheet. A single request can carry up to 9 reference images, 3 reference videos, and 3 reference audio clips, capped at 12 files combined. The per-type constraints are precise: reference videos run 2–15 seconds each with a combined ceiling of 15 seconds, in MP4 or MOV with H.264/H.265 encoding, up to 50MB each, at frame rates between 23.976 and 60fps. Reference audio accepts WAV or MP3 up to 15MB per clip, 2–15 seconds each and 15 seconds combined — and audio cannot travel alone: it must be accompanied by an image or video input.

For commercial work, this is the feature to test first. Reuters lists advertising, e-commerce, product design, and games as H3’s target use cases, and the reference system maps directly onto them: nine image slots cover a product’s angle set, a reference video can carry a motion style to imitate, and a free audio reference can pin a voice or music bed — all in one request instead of a multi-tool pipeline.

MiniMax publishes only unit rates — $0.13 per second at 2K and $0.09 per second at 768P — so we computed what complete clips cost across realistic reference-input scenarios. Two caveats before the table. First, the pricing page itself states that “768P is currently in closed beta” and tells users to contact sales before use, so the cheaper tier is not something you can simply switch on. Second, pay-as-you-go is the only mode H3 supports at launch — prepaid Hailuo consumer packages do not yet support H3.

Source: computed by Digital Applied from MiniMax’s published pay-as-you-go rates ($0.13/sec 2K, $0.09/sec 768P, first 5 reference images free then $0.04 each, reference audio free, reference video billed by duration at the output tier’s per-second rate). Note how reference video dominates the add-on math: a single 10-second reference clip costs two-thirds as much as the base generation itself at 2K, while nine reference images add just 16 cents. Reference audio, being free, is the obvious first lever for steering output without touching the budget.

On the competitive claim: MiniMax says “generating 2K video would cost less than one-third of mainstream rival products,” as reported by Reuters. That is a vendor statement with no named rival and no benchmark behind it — worth knowing, not worth repeating as settled fact.

MiniMax’s launch-day marketing leans hard on openness. The company’s official X account framed H3 with a four-part tagline ending in “Open Weights,” and OpenRouter’s hosted listing describes the model as “a lightweight, open-weights video generation model from MiniMax.” Here is what is actually true as of July 31: Reuters reports that MiniMax plans to release H3’s weights “within days” — a promise, not a release — and no MiniMax-H3 repository exists under the MiniMaxAI Hugging Face organization, where the newest entries remain M3 variants from June.

The gap matters because aggregator metadata is already muddying it. OpenRouter’s listing shows a “Released” date of July 29 — two days before the Reuters-reported launch — and its “open-weights” description is hosted-API listing copy, not evidence of a downloadable checkpoint. If your plans depend on self-hosting H3, the only signal that counts is an actual repository with actual weights and an actual license file. Until that exists, H3 is a closed, API-only model with an open-weights press release.

The same discipline applies to quality claims. No independent third-party benchmarks or arena scores exist for H3 as of this writing — no evals-leaderboard entry, no head-to-head study. Every cost and quality claim in circulation traces back to MiniMax’s own announcement, docs, or single-vendor write-ups. That is normal for a same-day launch; it just means the honest posture is to test on your own briefs before believing anything.

H3 did not launch into a quiet news cycle. ByteDance officially launched Seedance 2.5, its one-take video model, the very same day — July 31 — and Reuters explicitly frames H3 as intensifying MiniMax’s competition with ByteDance’s Seedance line and Kuaishou’s Kling in China’s fast-growing video-generation market. Kuaishou’s current flagship remains Kling 3.0, released in February 2026, with no newer Kling release in this window. For the wider landscape — including where Seedance and Google’s Gemini Omni sit — see our AI-video field guide.

Markets noticed. Secondary press reported that MiniMax’s Hong Kong-listed shares (00100.HK) jumped intraday on the H3 news, with reported gains ranging from roughly 13% to 17.4% of the share price depending on the source — the two figures circulate for the same session and the same ~HK$230 price level, so treat the percentage as reported-and-unreconciled rather than confirmed. The context: MiniMax, founded in 2022 and counted among China’s well-funded “AI tiger” startups, only listed in Hong Kong on January 9, 2026 — an IPO priced at HKD 165 that closed its first day up 109% on the offer price at HKD 345, raising roughly $619M. One Reuters variant adds that H3 was, as reported, engineered to work with Chinese-made semiconductor chips — a detail appearing in a single re-syndication, worth noting but not weight-bearing.

Here is our read on the trend: when no independent quality benchmarks exist — and none do, for H3 or for most same-week video launches — price becomes the only legible axis of competition, and vendors behave accordingly. MiniMax teased its H3-line video work at the World AI Conference around July 17, then launched two weeks later with a cost claim as the headline and per-second rates designed to undercut. A same-day rival launch, a stock pop on announcement, and a weights promise timed “within days” all point to the same conclusion: in this market, launch theater is strategy, and the durable facts — unit price, spec ceilings, reference-input systems — are what deserve your attention.

At $1.95 for a 15-second 2K clip with sound, the unit economics of video iteration change meaningfully — a hundred variant clips cost $195 in generation, before human review time. We break down the full ad-creative math in our companion analysis of what per-second pricing means for ad-creative economics. The decision tree by team type:

For ecommerce teams specifically, the reference system pairs naturally with the product-swap ad workflows we documented in our AI-video product-swap guide — the pattern of pinning a product with reference images and regenerating the surrounding creative. If you are weighing where AI video fits in a paid-media program, our paid media practice runs exactly these creative-economics evaluations, and our ecommerce team scopes catalog-scale video pipelines against real conversion data rather than launch-day spec sheets.

Looking forward: the next two weeks will tell you more than the launch did. If the weights ship as promised, H3 becomes the first serious open-weight 2K-with-audio video model and the self-hosting calculus changes overnight. If they slip, the “Open Weights” tagline becomes a data point about how much to trust the rest of the marketing. Either way, expect rivals to answer on price — per-second rates this visible tend to get matched — and expect the first independent benchmarks to land before any of these vendors’ claims settle into fact.