What Is MiniMax H3 (Hailuo 3.0)? The Open-Weight Multimodal Video Model, Explained

Viddi AI

ResterChed

Follow

August 1, 2026 · ~10 min read

Drop a product photo, a motion clip, and a voice recording into one prompt box, type a sentence, and MiniMax H3 returns a 2K video with native stereo audio — dialogue, effects, and room tone generated in the same pass, no separate audio stage. That is the pitch. Now the part that surprises people: the model most people call Hailuo 3.0 is officially MiniMax H3, it is marketed as open-weight, and as of one day after launch no weights have shipped — even as it ranks #1 in Video Editing, #2 in Text-to-Video, and #3 in Image-to-Video on the independent Artificial Analysis leaderboard. If you are choosing whether to build a video pipeline on it, here is what H3 actually is, what it costs, and when the open-weights promise becomes real.

TL;DR — MiniMax H3 (the official name of what most people call Hailuo 3.0) is the third generation of MiniMax's Hailuo video line, but it deliberately stops behaving like a video model. It reads text, images, video, and audio as one unified context, then generates a 4–15 s clip at 2K/24 fps with native stereo audio — no separate audio stage, no post-hoc upscaler. It supports up to 9 reference images, 3 reference video clips, and 3 reference audio clips per generation. On the Artificial Analysis leaderboards it ranks #1 in Video Editing, #2 in Text-to-Video, #3 in Image-to-Video. Weights are promised "in the coming days" under a planned MiniMax Community License (commercial use for organizations under $20M revenue, with attribution), but no Hugging Face model card exists yet as of August 1, 2026. Until the weights drop, you use the Hailuo app, the MiniMax Hub, the Open Platform API at $0.13/second of 2K video, or hosts like OpenRouter and Morphic.

This analysis is grounded in primary sources: the official MiniMax H3 announcement, the MiniMax Open Platform video-generation guide and V2 API reference, official pricing, Artificial Analysis leaderboard data, and hosting/spec docs from Morphic and OpenRouter. By the end you will know what H3 actually is, what it costs, what the open-weights situation really is, and whether to build on it now or wait.

What MiniMax H3 actually is (and why "Hailuo 3.0" is the same model)

MiniMax H3 is the official model name; Hailuo 3.0 — also written Hailuo 03 — is the widely used alias, after the Hailuo AI app it ships in. They are one model, the direct successor to Hailuo 2.3. MiniMax announced it on July 31, 2026 as "a general-purpose omni-modal generation model" that "jointly understand[s] multimodal contexts spanning text, images, video, and audio." Output is video with native stereo audio at up to 2K and 15 seconds. Morphic's model page confirms the naming — and warns it is not related to Kling O3, a Kuaishou model the "Hailuo 03" spelling is easily confused with.

The positioning is a deliberate break from the prior generations. MiniMax describes Hailuo 01 as building the system from scratch and Hailuo 02 as improving "architectural efficiency, data quality, and scale," then criticizes the industry's task silos — separate expert models for image editing, reference types, voice, SFX, music, and every video variant. H3's first design principle was "unifying and generalizing across tasks": image, video, and audio are trained jointly, voice/SFX/music are modeled as one audio stream, and reference-and-editing relationships are expressed in natural language instead of a fixed task set.

(Hailuo 2.3 figures per Morphic's H3-vs-2.3 comparison; H3 figures per the announcement and the API reference.)

The API reference labels the endpoint the "MiniMax video generation V2 (Hailuo-03) API" with model name MiniMax-H3 — one model, two names.

The architecture: one transformer, four engineered pillars

H3 is not tuned experts glued together; it is one multimodal model built on four named technologies, all described in the announcement:

Two details matter for practitioners. First, the 100K→4K token distillation is why H3 follows complex cross-modal instructions: the announcement's example prompt is "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3" — context becomes a compact language description of relationships. Second, MiniMax says it set aside the Hailuo-02 architecture as "unnecessary complexity." A full technical report is promised but not yet published, so treat this as the vendor's account, not a peer-reviewed spec.

What it outputs: native stereo audio, 2K, up to 15 seconds

The output contract separates H3 from every prior Hailuo model: one generation returns picture and sound together. Per the API reference:

Resolution, duration, and aspect ratios are from the API reference; the 24 fps output rate and the 1440 px short-edge definition are from Morphic and Artificial Analysis — the API reference does not state an output frame rate or pixel dimensions.

Note a spec discrepancy: Morphic and Artificial Analysis say 5–15 s, but the API enum allows 4–15 s — the API is ground truth. Morphic frames "2K" correctly: a 1440 px short edge, ~3.7 megapixels at 21:9. A 768p tier exists in the pricing table but is in closed beta.

Three generation modes (guide): text-to-video (prompt only; ratio required, not adaptive), first/last-frame image-to-video (0–2 images with first_frame/last_frame roles; ratio follows the input), and reference-to-video (prompt plus reference images/videos/audio — the multimodal mode that makes H3 interesting).

Reference images, video, and audio: how the multimodal context works in practice

Per the official guide and API reference:

• Images: ≤ 9 (JPG/JPEG/PNG/WEBP/HEIC/HEIF, ≤ 30 MB each)
• Videos: ≤ 3 clips (MP4/MOV, H.264/H.265, ≤ 50 MB each), 2–15 s per clip, ≤ 15 s total
• Audio: ≤ 3 clips (WAV/MP3, ≤ 15 MB each), 2–15 s per clip, ≤ 15 s total — audio cannot be sent alone; it must travel with an image or video
• Total: 12 files max; request body ≤ 64 MB (use public URLs, not Base64); prompt ≤ 7,000 characters
• Constraint: image-to-video and reference-to-video are mutually exclusive — no mixing first_frame/last_frame with reference_* roles in one request

In content terms, one generation can inherit a face from an image, a motion/camera move from a video, and a voice from an audio clip at once (Morphic: "a shot can inherit a face, a motion, and a voice at once"), alongside instruction-based editing and voice transfer.

Rule of thumb: think in reference sets, not single files — reuse the same 2–4 image + 1 video + 1 audio set across generations for character consistency; the first 5 reference images are free per generation.

Open weights: the promised release, and the license to watch

H3 is marketed as open-weight, but as of August 1, 2026 the weights have not shipped. The announcement says MiniMax plans "to open up the model weights in the coming days, subject to applicable laws and regulations," with hardware compatibility "a key consideration since the earliest stages of H3's design." On Hugging Face today, the MiniMaxAI organization page lists no H3 repository, and a model search for "minimax h3" returns zero results. No model card means no verified parameter count, no documented VRAM floor, no official benchmark table — everything about self-hosting hardware is still [INFERENCE].

The license is the other thing to watch. Artificial Analysis reports that MiniMax plans to release H3 "under the MiniMax Community License, which permits commercial use for organizations under $20M in revenue, subject to prominent attribution requirements" — and that if it ships, H3 "would become the strongest open weights video model, well ahead of the previous open weights leader LTX-2.3." That is a materially different deal from the MIT license MiniMax used on its earlier M-series checkpoints (M2, M2.5) — M2.7 was already the first release to break that streak, and its license change drew community criticism — read the final text when the repo lands.

Rule of thumb: plan for the weights under a Community License (commercial OK under $20M revenue + attribution), but do not commit to a self-hosted architecture until the repo, the technical report, and the license file all exist.

Using it today: the API, the pricing, and the access channels

Until the weights drop, the API is the only programmatic path, and it is a clean async contract (guide, V2 reference):

import os, time, requests
headers = {"Authorization": f"Bearer {os.environ['MINIMAX_API_KEY']}"}

# 1) Create an async task — multimodal content[] array with roles
payload = {
    "model": "MiniMax-H3",
    "content": [
        {"type": "text", "text": "Character speaks: Follow the wind, live free. Voice timbre follows reference audio 1."},
        {"type": "video_url", "video_url": {"url": "https://.../motion.mp4"}, "role": "reference_video"},
        {"type": "audio_url", "audio_url": {"url": "https://.../voice.mp3"}, "role": "reference_audio"},
    ],
    "resolution": "2K", "duration": 5, "ratio": "adaptive",
}
task_id = requests.post("https://api.minimax.io/v2/video_generation",
                        headers=headers, json=payload).json()["task_id"]

# 2) Poll status, then download the result URL directly
while True:
    time.sleep(10)
    task = requests.get(f"https://api.minimax.io/v2/query/video_generation/{task_id}",
                        headers=headers).json()["task"]
    if task["status"] == "succeeded":
        open("output.mp4", "wb").write(requests.get(task["content"]["url"]).content)
        break
    if task["status"] in ("failed", "cancelled", "expired"):
        raise RuntimeError(task.get("error"))

A callback_url option exists so you do not have to poll. Official access: the Hailuo AI app, the MiniMax Hub (three free H3 generations), and the Open Platform API. Third-party hosts verified at writing: OpenRouter (slug minimax/hailuo-3, "from $0.13/second") and Morphic.

Pricing (official, pay-as-you-go):

For scale: a 6-second 2K clip with audio costs $0.78, versus $0.28 for a 768p Hailuo 2.3 clip and $0.49 for its 1080p clip (official legacy pricing). One gotcha: the video subscription packages explicitly state "MiniMax H3 is not supported yet" — H3 is pay-as-you-go only.

Benchmarks and comparisons: what the data says so far

Cost is one axis; quality is the other. MiniMax published no benchmark table in the announcement — the technical report is pending. The only systematic third-party evaluation so far is Artificial Analysis, published the day of the launch: #1 in Video Editing (instruction-based editing, a capability "few models" offer), #2 in Text-to-Video, #3 in Image-to-Video (both with audio).

AA's per-minute cost comparison at 2K, measured on the same eval:

Two caveats before you cite these numbers. First, this is a single third-party arena run, not a peer-reviewed suite, and MiniMax's methodology is unpublished — treat the rankings as a prior, not a verdict, and re-run your own clips. Second, the "open weights leader" claim is prospective: conditional on the weights shipping under the planned license. Against the closed frontier the picture is clearer: H3 undercuts Kling 3.0 and Seedance 2.0 at 1080p by 2.5–3× while adding native audio and an editing mode those stacks would otherwise stitch from multiple tools.

Limitations and open questions

• Weights pending. No HF repo, no model card, no parameter count, no verified hardware floor — all self-hosting discussion is hypothetical.
• License ambiguity. The planned Community License (<$20M revenue + attribution) is a real constraint for funded teams; MiniMax's license history justifies reading the final text carefully.
• Eval gap. No official benchmark table, no technical report yet; one third-party leaderboard so far.
• Scope limits. 4–15 s per generation; audio references must piggyback on image/video input; i2v and r2v modes are mutually exclusive per request; no subscription tier, 768p closed beta.
• Vendor-acknowledged headroom. Next-H priorities: stronger multimodal understanding (with M-series LLM integration planned), model scale, higher resolution.

Responsible use: H3's voice transfer and reference-audio features are powerful — obtain consent before cloning anyone's voice or likeness in commercial work, and check output-tier terms before shipping client deliverables. Cost guardrail: model input video as billable (charged at the output rate); treat the 5 free reference images per generation as the default budget.

Decision framework: should you build on MiniMax H3?

Rules of thumb:

• Today vs. tomorrow: need output this week → API; need it on your GPUs → the weights — and don't pick infrastructure until the model card exists.
• Cost model: think per-minute ($7.80/min at 2K) and per-reference (5 images free, then $0.04); input video bills at the output rate.
• Consistency: reuse one reference set (face + motion + voice) across generations; don't re-describe characters in the prompt and expect identity to hold.
• Verify, don't trust: with no official eval, treat AA's #1/#2/#3 as a prior — run your own 3–5 shot test before committing a pipeline.

Watchlist for the next month: the weights release and license file (a community tracker keeps the status current), the full technical report, 768p leaving closed beta, and H3 landing in video subscription packages.

Frequently asked questions

Q: Is MiniMax H3 the same as Hailuo 3.0? Yes. H3 is the official model name; Hailuo 3.0 (also "Hailuo 03") is the alias from the Hailuo AI app, and the API itself labels the endpoint "Hailuo-03." It is unrelated to Kling O3 from Kuaishou.

Q: Is MiniMax H3 open weights? Planned, not shipped. The announcement says weights open "in the coming days, subject to applicable laws and regulations." As of August 1, 2026 there is no Hugging Face repo or model card; Artificial Analysis reports the planned license is the MiniMax Community License (commercial use for organizations under $20M revenue, with prominent attribution).

Q: What does MiniMax H3 cost? $0.13/second of 2K video ($7.80/min), per official pricing; a 768p tier at $0.09/s is in closed beta. Reference audio is free; the first 5 reference images are free, then $0.04 each; input video is billed by duration at the output tier rate.

Q: What are the input limits? ≤ 9 reference images, ≤ 3 reference video clips (2–15 s each, ≤ 15 s total), ≤ 3 reference audio clips (same limits, cannot be sent alone), 12 files total, prompt ≤ 7,000 characters (guide).

Q: Does H3 really generate audio? Yes — native stereo audio is generated with the picture: dialogue, effects, and room tone in the same pass, plus voice transfer from a reference recording (announcement, Morphic).

Q: How does H3 compare to Hailuo 2.3? Hailuo 2.3 tops out at 1080p/~10 s from a prompt or single image; H3 moves to 2K/15 s and adds mixed image-video-audio references, native stereo audio, voice transfer, and sentence-level editing (Morphic).

Q: Is MiniMax H3 open source? Not in the sense most developers mean. "Open weights" is planned — a downloadable checkpoint under the MiniMax Community License, which permits commercial use for organizations under $20M revenue with attribution — but the code, training data, and evaluation methodology are not open. It is also a different deal from the MIT license MiniMax used on its earlier M-series checkpoints (M2, M2.5), and the final license text does not exist yet; MiniMax's M2.7 license change is the reason to read the final file carefully before committing to self-hosting.

Q: When was MiniMax H3 released? Announced and available through the API on July 31, 2026. The open-weights release is promised "in the coming days"; as of August 1, 2026 there is no Hugging Face repository or model card.

Q: How does MiniMax H3 compare to Kling 3.0 and Dreamina Seedance 2.0? On the one independent Artificial Analysis comparison, H3 at 2K costs ~$7.80/minute versus ~$20.16 for Kling 3.0 at 1080p and ~$22.45 for Seedance 2.0 at 1080p — roughly 2.5–3× cheaper — while adding native stereo audio and an instruction-based editing mode those stacks would otherwise stitch from multiple tools. It is not a quality verdict: MiniMax published no official benchmarks, so run your own clips before choosing.

Key takeaways

MiniMax H3 is the most significant structural change in the Hailuo line since Hailuo 01: a single omni-modal model that replaces the task-siloed pipeline — T2I, I2V, editing, voice, SFX — with one unified context and one native-audio video output, priced to undercut the closed competition at 2K.

• H3 = Hailuo 3.0: one model, 4–15 s clips, 2K/24 fps, native stereo audio, three generation modes.
• Architecture bet: Omni Representation (language as bridge, ~100K→4K token distillation), H3-VAE (4× effective sequence length), Omni Transformer, in-context 2K regeneration.
• Practical now: API at $0.13/s (2K), 5 free reference images, async task contract — pay-as-you-go only, no subscription tier yet.
• Open-weights reality: promised "in the coming days" under a planned Community License; no HF repo or model card as of August 1, 2026 — audit before you commit to self-hosting.
• Eval: #1 video editing / #2 T2V / #3 I2V per Artificial Analysis — a single third-party run, not a verdict.

The smallest real test: take one reference set (a product image, a motion clip, a voice recording) and generate a single 5-second 2K clip through the Open Platform API — or claim the three free H3 tries on the MiniMax Hub — then compare against your current video stack on cost, audio quality, and how many tools you had to stitch together. If the weights drop while you are testing, read the license file on the MiniMaxAI Hugging Face org before you download the checkpoints — the license, not the architecture, is the detail most likely to change your decision.

Sources

• MiniMax H3 announcement (official blog, 2026-07-31)
• MiniMax Open Platform — Video Generation guide
• MiniMax Open Platform — Create Video Generation Task (V2 API reference)
• MiniMax Open Platform — Pay-as-you-go pricing
• MiniMax Open Platform — Video Packages pricing
• MiniMax Open Platform — H3 Feature Highlights
• MiniMax on Hugging Face (organization)
• MiniMax H3 on OpenRouter
• Morphic — MiniMax H3 model specs and FAQ
• Artificial Analysis — MiniMax H3 evaluation thread (X/Twitter)
• Decrypt — MiniMax M2.7 license change coverage