Black Forest Labs' FLUX 3 Max Generates 10-Second Videos 22x Faster
Black Forest Labs and fal shipped a post-trained FLUX 3 variant that renders a 10-second clip in 12 seconds, 22x faster than Seedance 2.5 at matched quality.
- Black Forest Labs and fal released FLUX 3 Max, a latency-optimized post-trained FLUX 3 Video variant
- Generates a 10-second 720p clip with native audio in about 12 seconds of inference time
- Claimed 22x faster than Seedance 2.5 at matched Elo, tied for #1 in image-to-video human eval
- Three endpoints: text-to-video, image-to-video, keyframes-to-video, up to 20-second clips at 24 fps
- Promotional pricing of $0.0425/sec at 720p and $0.0725/sec at 1080p through November 7
- Standard post-promo rates are $0.17/sec at 720p and $0.29/sec at 1080p on BFL
Black Forest Labs and fal have released FLUX 3 Max, a latency-tuned version of the FLUX 3 video model. The companies say it can generate a 10-second clip in about 12 seconds and runs 22 times faster than Seedance 2.5 at a matched Elo score in their human preference tests. Three fal endpoints are available, with launch pricing discounted by 75 percent.
Latency comes from post-training
FLUX 3 Max builds on FLUX 3 Video, Black Forest Labs’ multimodal foundation model for image, video, and audio generation. More than 95 percent of the base model’s training compute went toward video prediction using the company’s Self-Flow training method.
Fal adapted the checkpoint through post-training, which refines an existing model for specific behavior after its main training run. The work targeted lower latency, stronger prompt adherence, and higher visual quality. Max therefore serves as the speed-optimized FLUX 3 variant for product workflows that require rapid previews or multiple generations per request.
Read the benchmark narrowly
Black Forest Labs and fal report several performance results, all based on company-run tests rather than independent evaluations:
| Measure | Reported result | Scope |
|---|---|---|
| Generation time | About 12.45 seconds | 10-second clip |
| Sample inference time | 9.53 seconds | 10 seconds, 1280 × 704, 24 fps |
| Relative speed | 22 times faster than Seedance 2.5 | Matched Elo score |
| Preference ranking | Joint first place | Image-to-video evaluation |
Elo scores summarize pairwise human preferences: evaluators compare two outputs, and the model chosen more often receives the higher rating. The reported first-place tie applies only to image-to-video testing. The 22-times comparison also reflects a specific benchmark configuration, so production latency may vary with queueing, network overhead, concurrency, and endpoint settings.
Three endpoints cover the main workflows
Fal exposes FLUX 3 Max through three generation modes:
- Text-to-video generates a clip from a prompt.
- Image-to-video animates a source image and accepts an optional ending frame.
- Keyframes-to-video uses images placed at multiple points in the clip as guidance.
The documented schema keeps the main controls compact:
| Setting | Options |
|---|---|
| Duration | 5 to 20 seconds; 5 seconds by default |
| Aspect ratio | auto, 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, or 9:16 |
| Audio | Enabled by default with generate_audio |
| Reasoning level | low or high, controlling prompt expansion |
| Output | 720p video at 24 fps with synchronized audio |
Prompts can include dialogue, environmental audio, and sound-effect cues. Black Forest Labs also claims support for realistic motion, synchronized speech, and unusual visual concepts, although those capabilities require prompt-level testing for each application.
The broader FLUX 3 Video model supports clips up to 20 seconds, resolutions up to 4K, multilingual speech, lip synchronization, effects, and ambience in one request. The current Max endpoints document a 720p output ceiling, so developers should not assume that every base-model capability carries over.
Launch pricing favors high-volume testing
Fal’s October 15 to November 7 promotion reduces usage prices by 75 percent. Billing is based on generated video duration:
| Resolution | Promotional rate | Standard rate | 10-second promotional cost | 10-second standard cost |
|---|---|---|---|---|
| 720p | $0.0425 per second | $0.17 per second | $0.425 | $1.70 |
| 1080p | $0.0725 per second | $0.29 per second | $0.725 | $2.90 |
Fal publishes a 1080p rate even though the current endpoint description specifies 720p output. Teams requiring 1080p should confirm endpoint availability and billing behavior before integrating it.
At standard pricing, Max matches Black Forest Labs’ listed rates for standard FLUX 3 Video at HD and Full HD. The broader range costs $0.06 per second for Draft HD, $0.17 for HD, $0.29 for Full HD, $0.40 for QHD, and $0.80 for UHD. Max’s cost advantage is therefore concentrated in the promotional period; its longer-term case rests on latency.
Where 12-second generation fits
Video models that take 30 seconds or several minutes per clip are difficult to place inside interactive workflows. Generation near the duration of the resulting clip makes several product patterns more practical:
- Agent workflows that generate and rank several candidates
- Editors that preview motion after prompt or keyframe changes
- Advertising pipelines that produce many variants from one brief
- Short-form content tools with strict per-clip budgets
- Keyframe-guided storyboards assembled into longer sequences
Actual suitability depends on more than raw inference time. A production evaluation should measure end-to-end latency, queue time, temporal consistency, prompt adherence, audio quality, lip synchronization, failure rates, concurrency limits, and cost across representative prompts.
Benchmark the full product path
Independent results have yet to validate the companies’ quality rankings or the 22-times speed claim. The comparison covers Seedance 2.5 at a matched Elo score, while the first-place result covers image-to-video only. Teams evaluating FLUX 3 Max should also account for the documented 720p ceiling and the higher rates that apply after the launch promotion.
Comments
