A ringed planet appearing beyond a flooded celestial observatory in an H3 SuperFast generation
DDG field test · H3 SuperFast

Twelve seconds.
Generated in 4.21.

H3's architecture is built for audiovisual efficiency. We tested whether a performance-tuned production deployment could make the “SuperFast” name survive contact with real prompts.
9 original generations · 6 aspect ratios · text and image

H3 SuperFast is the fastest high-quality video model we have measured on this route. Every clip in our evaluation requested 12 seconds of 768p video with native stereo audio. Backend metadata reported GPU denoising times between 4.21 seconds and 12.09 seconds—from real-time to almost three times faster than playback.

4.21sFastest backend inference for a 12-second clip
12.09sSlowest backend inference in our nine-clip run
24 fpsMeasured across every downloaded output
80% offDDG launch price against the sustainable full rate

This Is Not “Fast for AI Video.” It Is Fast.

The production deployment is tuned for stronger prompt adherence, aesthetics, and throughput. The important part is not the claim—it is the result: all nine of our twelve-second requests completed their measured denoising pass at approximately real-time speed or faster.

A 12.256-second square video finished its GPU denoising pass in 4.213 seconds.

That is 2.91× faster than the finished clip's playback length. Across all nine requests, reported inference ranged from 4.21 to 12.09 seconds. End-to-end wait plus download was longer—14.56 to 55.75 seconds, with a 34.16-second median—because it includes queueing, quality prompt expansion, encoding, transfer, and our concurrent workload.

The distinction matters. The backend's timings.inference field measures the GPU denoising pass, not the stopwatch time from click to a playable MP4. We therefore call H3 SuperFast a near-real-time generation engine, not a guaranteed five-second delivery service. Network and queue conditions still exist.

Seven Original Clips, No Vendor Demo Reel

We submitted six text prompts across the complete ratio set, then animated three synthetic source frames. Every output below is the raw downloaded MP4: 768p tier, 12 requested seconds, quality prompt expansion, native stereo audio, and no post-production edit.

01
1536×67221:910.78s inference

The Sunken Observatory Wakes

Our strongest text result. A dark flooded hall progresses through a turning brass orrery, constellation light, schooling fish, an opening roof, and a ringed planet. The ultra-wide frame earns its width instead of merely stretching the scene.

02
Image to video16:912.05s inference

The Glass Train Crosses a Living Desert

The best source-preservation test. Reflective locomotive geometry, botanical carriages, koi, dust, and prismatic light stay coherent through a long side-tracking move.

Synthetic glass train source frameStarting frame
A synthetic frame from our earlier Wan 3.0 evaluation. H3 SuperFast preserved the train's core design while creating new movement and environmental detail.
03
768×7681:14.21s inference

Clockwork Garden Bloom

The speed champion and one of the most controlled transformations: frost retreats, buds appear, orchids open in a radial sequence, and the brass clock remains legible at the center.

04
1344×76816:910.67s inference

The City Folds Into Dawn

A courier and electric skiff remain remarkably stable as blue hour turns to sunrise and the architecture becomes progressively more paper-like. The literal city-block fold is simplified, but the shot never loses its subject.

05
Image to video1:14.65s inference

The Celestial Whale Creates a Galaxy

The star-field skin and cloud world survive the animation while a galaxy and orbiting moons form between the whale's fins. The anatomy becomes rounder head-on, but the source identity remains unmistakable.

Synthetic celestial whale source frameStarting frame
The first frame is shown here so source fidelity can be judged directly rather than taken on faith.
06
768×13449:1611.98s inference

Gravity Circus Above the Clouds

The portrait frame becomes a stage for height: a crimson aerialist moves from a needle tower into vertical running and hoop action. Costume and subject remain readable through the camera's climb.

07
1024×7684:36.70s inference

Storm in a Teacup

A tactile miniature lighthouse, blueberries, and a human-scale intervention sell the tiny-world illusion. The requested lightning-to-snow beat is only partially visible—proof that speed does not remove event-coverage limits.

What We Actually Measured

EvidenceMeasured resultInterpretation
Output success9 of 9 requests completedSix text videos and three image videos; no retry was needed.
Backend inference4.213–12.093s; median 10.667sEvery 12-second request was denoised at approximately real-time speed or faster.
Wait + download14.563–55.751s; median 34.157sReal user latency also includes queue, prompt expansion, encoding, and transfer.
Delivered media12.256s each; 24 fps; H.264 + AAC stereoAll nine outputs contained both video and audio streams.
Ratio coverage21:9, 16:9, 4:3, 1:1, 3:4, 9:16The text endpoint composes natively across the complete advertised set.
Visual review4.0–4.8 / 5 from six sampled framesStrong coherence overall; pseudo-text and omitted secondary beats remain failure modes.
Audio scope: we verified that every MP4 contains 32 kHz AAC stereo audio. Contact sheets cannot judge sound quality, lip sync, or smooth motion between sampled frames, so those qualities are not scored here.

How Can a 33B Video Model Be This Fast?

The implementation details of the production serving stack are not public. MiniMax has, however, documented the architecture of the H3 foundation underneath it.[1] That lets us separate facts from responsible engineering inference.

What we know

  • Compressed video latents: H3's visual VAE compresses space by 16× and time by 4×, reducing the sequence the transformer must process.[1]
  • One joint audiovisual model: a 33B dense, single-stream Omni Transformer predicts video and audio latents together instead of running a separate soundtrack pipeline.[1]
  • Cacheable modulation: roughly 13B parameters live in AdaLN branches whose outputs can be precomputed and cached for inference-only deployment.[1]
  • CFG-distilled checkpoints: the released H3 weights have classifier-free guidance distilled into the model, avoiding the conventional doubled conditional/unconditional pass.[1]
  • Sparse-attention design: H3 was trained with native sparse attention for long sequences, although MiniMax says the first public implementation shipped with full attention only.[1]

What we infer

  • The production stack likely keeps the model resident on high-bandwidth accelerator pools, avoiding the aggressive CPU offload that makes local H3 runs take minutes.
  • The custom stack plausibly combines fused attention/normalization kernels, precomputed AdaLN modulation, compilation, and sequence parallelism.
  • The serving deployment may also use a tuned scheduler or fewer effective denoising steps than the general public checkpoint.
  • The serving stack may have access to MiniMax's native sparse-attention path, but that has not been confirmed. We will not present it as fact.
  • Warm batching and aggressive GPU utilization can explain throughput, but they do not guarantee identical click-to-download latency when the queue is busy.

MiniMax's own architecture story supports the direction of these assumptions. The company says its redesigned tokenizer produces a fourfold gain in effective sequence length, while separating understanding and generation workloads raised training throughput by nearly 30%.[2] Those are vendor-reported training and architecture figures—not proof of the production runtime recipe—but they show that efficiency was designed into H3 before the serving work began.

80% Off H3 SuperFast at Launch

H3 SuperFast launches at just one-fifth of its normal DDG energy price. That means more ideas, more variations, and more ambitious shots for the same energy—whether you are exploring fast 480p drafts or finishing a 768p scene with native stereo audio.

Turbo · 480pNormally 20 energy/s
Launch: only 4 energy/s
Pro · 768pNormally 31 energy/s
Launch: only 6 energy/s

The discount applies automatically to every H3 SuperFast generation. No coupon and no special plan—just choose your tier and create at the launch rate.

Where H3 SuperFast Fits

  • Choose H3 SuperFast when iteration speed, cinematic motion, native audio, or rapid image animation matters most.
  • Choose Wan 3.0 when you need a single 20–30 second generation or 1080p output.
  • Use 480p for prompt exploration and variant generation; move the winning concept to 768p.
  • Keep prompts staged: one main subject, one camera journey, and two or three visible beats. Fast generation does not make a twelve-event checklist reliable.

Make the next version before the idea cools.

Start with a five-second 480p draft, compare several directions, then render the strongest shot at 768p—with native stereo audio included.

Open H3 SuperFast →

Sources and test notes

  1. MiniMax — official H3 open-model architecture, components, and deployment notes
  2. MiniMax — H3 technical choices, tokenizer efficiency, and in-context regeneration

DDG internal evaluation: nine completed requests, complete request/result metadata, ffprobe validation, six-frame contact sheets, and a scored visual review completed August 27, 2026. “H3 SuperFast” is DDG's product name for this performance-tuned H3 deployment.

Comments (0)