Model reference · Generator
MiniMax H3 Max is a post-trained version of MiniMax H3 developed by fal Research, rather than a separate “Max” edition released by MiniMax. fal starts from the open-weight H3 base, adds post-training focused on prompt adherence and visual quality, and serves the result through an inference stack co-optimized for high throughput.
The current H3 Max endpoints focus on fast text-to-video and image-to-video generation at 480P or 768P with synchronized audio. Standard MiniMax H3 remains the better fit when you need 2K output or its broader reference-to-video workflows.
H3 Max by fal Research
Create a video from text or animate a first frame with synchronized audio.
Key facts
| Attribute | Current fact |
|---|---|
| Developer | fal Research |
| Base model | MiniMax H3 |
| Release | August 27, 2026 |
| Type | Post-trained H3 variant |
| Public launch workflows | Text to Video / Image to Video |
| Resolution | 480P / 768P |
| Default resolution | 768P |
| Duration | 5–15 seconds |
| Text-to-video aspect ratios | 21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 |
| Image-to-video composition | Follows first-frame aspect ratio |
| Last-frame control | Optional end image |
| Audio | Synchronized/native audiovisual generation |
| Prompt expansion | Disabled / Balanced / Quality |
| Primary positioning | Speed + prompt adherence + aesthetics |
MiniMax H3 Max is fal Research's post-trained variant of the open-weight MiniMax H3 model. fal added substantial post-training data with an emphasis on prompt adherence and visual quality, while its inference team co-optimized the serving stack for much higher throughput.
H3 Max should therefore be understood as a derived H3 model and serving system, not simply as a higher-resolution or premium-named version of MiniMax H3.
fal starts with MiniMax H3 open weights, adds substantial new training data, and evaluates checkpoints for prompt understanding and aesthetics. Its launch materials also describe reinforcement-learning work with verifiable rewards.
fal developed the serving stack alongside the model, applying custom diffusion inference and kernel work while preserving the post-training quality gains. Published throughput therefore reflects the combined model and system.
H3 Max's speed is not only a model-weight story. fal describes H3 Max as the result of post-training and inference-system co-design, so its published latency reflects both the derived model and the serving stack.
Comparison
| Attribute | H3 Max | MiniMax H3 |
|---|---|---|
| Developer | fal Research | MiniMax |
| Relationship | Post-trained H3 derivative | Base/original model family |
| Primary focus | Speed + prompt adherence + aesthetics | Broader general-purpose multimodal workflows |
| Resolution | 480P / 768P | Up to 2K |
| Duration | 5–15 sec | 5–15 sec on current fal H3 endpoints |
| Native/synchronized audio | Yes | Yes |
| Text to Video | Yes | Yes |
| Image to Video | Yes | Yes |
| First + Last Frame | Yes | Yes |
| Reference to Video | No live public Max endpoint verified | Yes |
| Rich image/video/audio references | No equivalent current public Max workflow verified | Yes |
| Separate public H3 Max weights | No separate release currently stated | H3 open weights available |
| Best fit | Fast iteration / throughput | 2K / richer multimodal reference control |
H3 Max is not simply H3 with higher specifications. Its main advantages are fal's post-training and serving speed, while standard MiniMax H3 retains 2K output and broader reference-to-video workflows.
H3 Max is primarily a speed-and-adherence option, not a universal replacement for MiniMax H3.
Evidence
fal reports that H3 Max can generate a 5-second 768P video in under three seconds on its optimized serving stack. fal also reports roughly 35× the throughput of the official H3 endpoint in its launch comparison.
Vendor-reported launch benchmark
| Claim | Evidence | Type |
|---|---|---|
| 5-second clip in under ~3 seconds | fal launch benchmark | Vendor-reported |
| ~35× official H3 endpoint throughput | fal launch benchmark | Vendor-reported |
| Design Arena ranking | External leaderboard cited by fal | Leaderboard snapshot |
| Artificial Analysis ranking | External leaderboard cited by fal | Leaderboard snapshot |
Leaderboard positions and provider benchmarks are snapshots, not permanent model specifications.
Resolution
480P / 768P
Duration
5–15 seconds
T2V ratios
21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16
I2V input
First frame; optional last frame
I2V composition
Follows first image
Prompt expansion
Disabled · Balanced · Quality
Audio
Synchronized audio
Safety
Provider safety checking stays enabled
| Mode | Behavior |
|---|---|
| Disabled | Send the original prompt without fal's expansion step |
| Balanced | Fast prompt expansion; default/recommended |
| Quality | Spend additional time producing a richer expanded prompt |
The expanded prompt is a preprocessing layer exposed by fal's H3 Max API. It is not the H3 Max model itself.
Learn the MiniMax H3 prompt structureProvider price snapshot
As checked August 29, 2026, fal lists promotional launch rates of $0.025/second at 480P and $0.04/second at 768P. fal states that the promotion ends September 1, after which list rates are $0.05/second and $0.08/second respectively.
Estimated minimax3.org cost
The generator quote uses minimax3.org credits per output second. These retail credits are configured independently from fal's provider pricing and the current estimate shown before generation is authoritative.
TEXT TO VIDEO
minimax/h3-max/text-to-videoprompt · duration · resolution · prompt_expansion_mode · aspect_ratio · seed
IMAGE TO VIDEO
minimax/h3-max/image-to-videoprompt · duration · resolution · prompt_expansion_mode · image_url · end_image_url · seed
Relevant output fields are video, expanded_prompt, and timings.inference. Provider keys must stay on the server.
MiniMax H3 itself has open weights. H3 Max is a fal Research post-trained derivative, but fal's current launch materials do not announce a separate downloadable H3 Max checkpoint. Users currently access H3 Max through fal's hosted product and API.
If separate H3 Max weights are released later, this section should be updated.
H3 Max is fal Research's post-trained variant of the open-weight MiniMax H3 model, co-optimized with fal's inference stack for speed, prompt adherence and visual quality.
No. MiniMax developed the H3 base model. fal Research developed the H3 Max post-trained derivative and serving system.
It depends on the workflow. H3 Max prioritizes speed and prompt adherence at up to 768P; standard H3 is the stronger fit for 2K and broader multimodal reference workflows.
fal reports a 5-second video in under roughly three seconds on its optimized serving stack. This is a vendor-reported launch benchmark, not a guaranteed end-to-end request time.
The current public H3 Max endpoints expose 480P and 768P output.
Current fal product documentation supports single generations from 5 to 15 seconds.
Yes. fal describes H3 Max as retaining H3’s natively synchronized audio and video capability.
Yes. Its image-to-video endpoint accepts a first-frame image and derives the output composition from that image.
Yes. Add an optional end_image_url to the image-to-video endpoint to control the final frame.
It is a fal API preprocessing control that rewrites the prompt before generation. Modes are disabled, balanced and quality.
The current H3 Max endpoints do not expose 2K. Standard MiniMax H3 supports 2K workflows.
MiniMax H3 has open weights. fal’s current H3 Max launch materials do not announce a separate downloadable H3 Max checkpoint.
No separate H3 Max checkpoint is currently announced. H3 Max is currently accessed through fal’s hosted product and API.
fal exposes minimax/h3-max/text-to-video and minimax/h3-max/image-to-video through its queued model API.
Source ledger
| Claim | Source |
|---|---|
| H3 Max is fal's post-trained H3 derivative | fal H3 Max launch article |
| Post-training + inference co-design | fal H3 Max launch article |
| 480P / 768P and T2V ratios | fal T2V API schema |
| 5–15 second support | fal H3 Max product documentation |
| I2V follows first image aspect ratio | fal I2V API schema |
| Optional final frame | fal I2V API schema |
| Prompt expansion, expanded_prompt, timings.inference | fal API output schema |
| Standard H3 supports 2K / richer references | fal standard H3 documentation |
Last verified: August 29, 2026.