Evidence review · 2026

H3 Max Turbo review: verified public evidence

This evidence review is based on official records, provider data, independent public tests, and external benchmarks. It does not claim FrameSignal hands-on testing.

Updated Sep 23, 2026Model brief
External evidence summary. A FrameSignal hands-on review has not been completed. The source reports below are also included in the model details.

Current evidence summary

H3 Max Turbo is fal Research’s post-trained MiniMax H3 variant, exposed through separate text-to-video and image-to-video endpoint families. The canonical fal routes support 5–15 second clips and 480P, 768P, or 1080P output, with native generated audio. Creative Claw’s “H3 Max Fast” label maps to the Turbo route and is not a separate model. An exact canonical release date is not verified; the first dated fal-owned availability reference found in this research was September 9, 2026. This record keeps Turbo’s provider pricing, evidence, and benchmarks separate from H3 Max. The source-backed use cases currently point to fast hosted text-to-video and image-to-video iteration, native-audio video generation, developers comparing usd-per-second, provider credits, and creative units without conflating them.

Public benchmark snapshots

Each result stays attached to its task, date, sample count, and reported uncertainty. Developer backend timings are kept separate from end-to-end wait tests.

V-Benchmark v2 · Sep 14, 2026

H3 Max Turbo

independent video generation benchmarkMegaton Index73.09 score 0 100 · rank 9
independent video generation benchmarkprompt adherence87.64 score 0 100 · n=38
independent video generation benchmarkscene consistency84.59 score 0 100
independent video generation benchmarkphysics50.59 score 0 100
Open benchmark source ↗
H3 Max family inference timing · Sep 8, 2026

H3 Max Turbo

developer reported backend inferenceinference time 5s 768p1.54 seconds
developer reported backend inferenceinference time 15s 768p8.44 seconds
Open benchmark source ↗
Fastest AI Video Models September 2026 · Sep 1, 2026

H3 Max Turbo

independent endpoint speed testmedian end to end 5s 768p3.8 seconds · rank 1
Open benchmark source ↗

External tests and provider examples

These are reports from the named publishers and providers; FrameSignal did not run these generations.

Creative Claw · Itay Dressler / Creative Claw · reliability A

controlled same frame product i2v

Reported metric: platform charged cost — 0.4 USD

Creative Claw published a controlled same frame product i2v example. The displayed H3 Max Fast label maps to the H3 Max Turbo route; its $0.40 charge is Creative Claw platform pricing, not fal API pricing.

Reported result: Activation beats were visible and opening composition was preserved; a new vent detail appeared late in the orbit and audio became mostly quiet after the opening sound event.

Prompt: Starting from the supplied frame, the compact coral-orange desktop media console powers on; four indicator lights brighten in sequence, central dial turns slightly, warm light travels across the front edge; slow 15-degree move right plus subtle push-in; quiet room tone, electronic chime and dial click; preserve exact geometry/color/controls with no added objects or deformation. Full exact prompt at source.

Settings: same reference frame: true · duration seconds: 5 · resolution: 768p · aspect ratio: 16:9 · single output: true · rerolls: 0 · provider label: H3 Max Fast · actual model route: video/minimax-h3-max-turbo

Open source ↗
Voyager · Anvisha Pai / Voyager · reliability A

text to video product turntable

Reported metric: submit to result wait — 6 seconds

Voyager published a text to video product turntable example. The reported time is submit-to-result wait, not backend inference time.

Reported result: Estimated $0.10 under the Sep promotion; sampled output gained a spout-like lip and handle, showing product-identity drift.

Prompt: A matte black ceramic pour-over coffee dripper rotating slowly on a white turntable, soft studio light, seamless white background, product video, no text.

Settings: endpoint: minimax/h3-max-turbo/text-to-video · prompt expansion mode: disabled · aspect ratio: 16:9 · resolution: 768P · duration seconds: 5

Open source ↗
Voyager · Anvisha Pai / Voyager · reliability A

text to video talking head

Reported metric: submit to result wait — 6 seconds

Voyager published a text to video talking head example. The reported time is submit-to-result wait, not backend inference time.

Reported result: Returned one-shot talking-head output with generated voice; audio was present but Voyager did not score lip-sync by ear.

Prompt: A woman in her 40s with short grey hair stands in a bright kitchen, looks at camera and says 'Good morning, the coffee is ready', natural window light, handheld feel, she smiles at the end.

Settings: endpoint: minimax/h3-max-turbo/text-to-video · prompt expansion mode: disabled · aspect ratio: 16:9 · resolution: 768P · duration seconds: 5

Open source ↗
Voyager · Anvisha Pai / Voyager · reliability A

text fidelity title card

Reported metric: submit to result wait — 6 seconds

Voyager published a text fidelity title card example. The reported time is submit-to-result wait, not backend inference time.

Reported result: Text was incomplete during entrance but resolved into readable lettering later.

Prompt: A title card reading 'MODA SUMMER SESSIONS' animates in over slow-motion festival footage; smaller line 'Friday 12 June' fades in beneath it.

Settings: endpoint: minimax/h3-max-turbo/text-to-video · resolution: 768P · duration seconds: 5 · prompt expansion mode: disabled

Open source ↗
Voyager · Anvisha Pai / Voyager · reliability A

image to video product

Reported metric: submit to result wait — 6 seconds

Voyager published a image to video product example. The reported time is submit-to-result wait, not backend inference time.

Reported result: Returned 768x768; model filled the source dripper with coffee instead of showing a filter draining, a concrete semantic failure.

Prompt: The coffee dripper slowly fills as coffee drips through it, steam rising, the camera drifting gently around it, the dripper itself unchanged.

Settings: endpoint: minimax/h3-max-turbo/image-to-video · resolution: 768P · duration seconds: 5 · prompt expansion mode: disabled · source image: true

Open source ↗
Runware · Runware · reliability A

text to video provider example

Reported metric: raw generation charge — 0.16 USD

Runware published a text to video provider example example. This is an external/provider report, not a FrameSignal generation.

Reported result: Provider displays approximately 44 seconds end-to-end for the published example.

Prompt: Semiconductor wafer transfer B-roll; full prompt and request at source.

Settings: resolution: 768p · workflow: text-to-video

Open source ↗
fal · fal · reliability B

same prompt speed tier comparison

Reported metric: inference time example — 1.61 seconds

fal published a same prompt speed tier comparison example. This is a developer-authored test, not an independent benchmark.

Reported result: fal reports H3 Max Turbo as the fastest endpoint in this developer-authored comparison; treat as vendor test, not independent benchmark.

Prompt: Same vertical skincare UGC prompt across five speed-tier endpoints; full prompt/settings at source.

Settings: same prompt: true · comparison set: ["H3 Max Turbo","H3 Max","Kling V3 Turbo Pro","Seedance 2.0 Mini","LTX-2.5 Fast"]

Open source ↗

What official evidence suggests

Promising areas

  • Separate official fal T2V and I2V endpoints with 5–15 second duration and three output tiers
  • fal examples show generated sound; Layer separately documents stereo audio
  • Turbo-specific Megaton benchmark and process-rich third-party tests are available
  • “H3 Max Fast” is resolved to the Turbo route rather than duplicated as another model

Verify before paying

  • No exact canonical release date has been verified
  • A dedicated Turbo reference-to-video or general video-editing endpoint remains unverified
  • Promo rates are time-limited and provider prices use incompatible units
  • Backend inference timings and end-to-end waits are different metrics; the evidence is external, not FrameSignal testing