On October 19, 2026, fal Research released FLUX 3 Max, a video generation AI model developed jointly with Black Forest Labs (BFL). The model is built by further training BFL's foundation model FLUX 3 and optimizing its inference. According to data published by fal, generating a 10-second video now takes 12.45 seconds, down from 126.44 seconds. fal also says it is about 22 times faster than the competing Seedance 2.5. The company aims to improve video quality and prompt adherence while sharply reducing the waiting time involved in repeated trial and error during video production.
However, the generation time fal published does not necessarily match the time users actually wait before receiving a video. The inference time returned by the API does not include pre- and post-processing, and processing speed also varies with generation settings. In addition, the model's top-tier placement in the quality ranking applies to image-to-video generation, and the same result does not necessarily hold for text-to-video. How much benefit you get in a real production environment should be judged with these conditions in mind.
Additional training and inference optimization cut generation time to about one-tenth
FLUX 3 Max is based on FLUX 3, the foundation model developed by BFL, with additional training by fal Research in collaboration with BFL. According to fal's official description, the additional training improved visual fidelity, motion consistency, and prompt adherence, and also targeted better quality when generating video from images. By building on the visual expression that the existing model has already learned and training it further, the aim is to raise the quality of the output.
Alongside the additional training, fal Research also optimized inference for the trained model. By revisiting how the model is run, it shortened the time needed to produce video and increased the number of requests that can be processed in a given period. It also used its own inference engine to tune the systems that support model execution and delivery. In other words, the speedup is said to come not from additional training alone but from a combination of changes to how the model runs and improvements to the surrounding systems. How much each step contributed to the shorter generation time has not been disclosed.
According to the comparison table fal published, the average time to generate a 10-second video was 126.44 seconds for the earlier FLUX 3 and 12.45 seconds for FLUX 3 Max, a speedup of about 10.2 times. The same table lists Seedance 2.5 at 277.80 seconds, which makes FLUX 3 Max about 22.3 times faster. All the compared videos were 10 seconds long, with a 16:9 aspect ratio and a resolution of 720p or 768p, and the figures are average generation times under the conditions fal published.
This kind of speedup is likely to be most useful in video production, where generation is repeated many times. For example, when changing camera movement or how a subject is staged, a shorter gap between checking a result and trying the next instruction makes it easier to adjust the footage on the spot. If a wait of several minutes drops to a dozen or so seconds, it could affect how the production process is run. That said, the published averages were measured under specific conditions, and the same speed is not guaranteed when servers are busy or with different settings.
The published 12.45 seconds and the API's inference time are not necessarily the same
When assessing FLUX 3 Max's speed, it is necessary to distinguish the "12.45 seconds" in fal's comparison table from the inference time returned by the API. The API specification for text-to-video defines timings.inference as the time, in seconds, that the model spent denoising the video. In video generation AI, the core process is building up footage while removing noise step by step, but other processing is also required before the finished video reaches the user.
fal's API specification states explicitly that timings.inference "does not include prompt expansion, decoding, or upload." Prompt expansion is a process that supplements the instruction that was entered, and decoding converts the model's computed output into a format that can be handled as video. The time to upload the generated video is also excluded from inference time. Therefore, the inference time returned by the API alone does not reveal the total time before a user receives the video.
The differences among the times that can be confirmed from public materials are summarized below.
| Type of time | What it measures / what can be confirmed | Usage note |
|---|---|---|
| 12.45 seconds in fal's comparison table | Average time to generate a 10-second, 16:9, 720p/768p video | Usable for comparing models listed in the same table, but its correspondence to API inference time is not stated |
API timings.inference |
Computation time the model spent denoising the video | Excludes prompt expansion, decoding, upload, and similar steps |
| Time from sending a request to retrieving the video | Time from the user sending a request to the API until receiving the generated video | Must be measured in your actual usage environment and generation settings |
As this shows, what is being timed is not necessarily the same in fal's comparison table and in the API specification. It has not been made clear whether 12.45 seconds corresponds directly to the timings.inference value, so it is not appropriate to equate the two, or to guess at the unknown pre- and post-processing time and add it on.
Another setting that affects generation speed is reasoning_level. It can be set to low or high, and the default is high. fal explains that the more computation devoted to processing, the longer generation takes, but it has not said how much quality or speed differ between the two settings. For that reason, the published 12.45 seconds should not be assumed to be reproducible under every setting.
Note also that fal's model introduction page says a 10-second video is generated in "about 11 seconds" and is "more than 25 times faster" than Seedance 2.5. These figures differ from the 12.45 seconds and roughly 22 times on the dedicated page, and it is unclear whether they were measured under the same conditions. When comparing, check the measurement conditions and the source of each figure rather than mixing them up.
In image-to-video quality evaluation, it is statistically tied with the leader
Beyond generation speed, fal also claims FLUX 3 Max has reached a level of video quality on par with competing models. Its basis is "AA-Video-I2V v1.1," from Artificial Analysis, which benchmarks AI model performance. In this evaluation, videos with audio are generated from a source image, and people compare two generated results and choose which they prefer. The results are converted into relative Elo scores, from which each model's rating is calculated.
In Artificial Analysis's leaderboard, FLUX 3 Max placed second with 1138 Elo, while MiniMax H3 led with 1140 Elo. The ratings of the top three models are as follows.
| Model | Elo score | 95% confidence interval |
|---|---|---|
| MiniMax H3 (768p) | 1140 | 1124–1156 |
| FLUX 3 Max (720p) | 1138 | 1123–1153 |
| Dreamina Seedance 2.5 | 1138 | 1121–1155 |
MiniMax H3 is slightly ahead in the Elo point estimates, but the 95% confidence intervals of the three models overlap substantially. A confidence interval expresses the uncertainty of an estimate, so these figures alone cannot establish a small score gap between models as a decisive performance difference. fal describes FLUX 3 Max as "tied for first" on the grounds that the gap to the top models is not clear. Still, overlapping confidence intervals do not prove that quality is completely equal across models.
The provisional rating on fal's dedicated page gives FLUX 3 Max 1139 and MiniMax H3 1141, slightly different from the figures on Artificial Analysis's leaderboard. Ratings fluctuate with the number of votes and the time of tabulation, so it cannot be concluded that the models' performance itself changed by comparing figures from different points in time. VentureBeat's October 19 article also introduces FLUX 3 Max's speedup and notes that the confidence intervals of the top models overlap.
Another important point is that this evaluation covers image-to-video, which generates video from an image. FLUX 3 Max can also generate video from text alone, but this ranking is not a direct evaluation of that capability. For example, adding motion while preserving the appearance of a person or product in a source image demands different abilities from composing an entire scene using text alone. When comparing quality, check evaluation results that match the generation method you will actually use.
Supports text, images, and keyframes; 75% off through November 7
In addition to text-to-video, FLUX 3 Max offers APIs for image-to-video and keyframes-to-video, which lets you specify multiple keyframes. Image-to-video lets you specify a starting image for the video and, optionally, an ending image. Keyframes-to-video lets you place up to 10 images at distinct, non-overlapping frame positions, which also allows you to specify scenes in the middle of the video.
For example, if you want to add motion to a product photo, you can specify a starting image, and if the final composition is already decided, you can also set an ending image. If you want to insert a particular scene partway through the video, using multiple keyframes is one option. Note that fal's dedicated page includes a passage, as of October 18, describing two input methods, but the currently published API specification also includes the keyframe method. When checking available features, refer to the latest specification for each API.
Videos can be 5 to 20 seconds long, specified in 1-second increments. The default is 5 seconds, and audio generation is enabled by default but can be turned off if needed. Aspect ratios include not only landscape 16:9 but also portrait 9:16 and square 1:1. The default resolution is 720p, and 1080p output can also be selected. However, 1080p is produced by upscaling, so this difference should be kept in mind when comparing quality with models that generate native 1080p.
On pricing, fal is running a 75% discount campaign through November 7, 2026. Billing is based on the length of the generated video, so the price is determined by the seconds of output video rather than the time spent on inference. Text-to-video and image-to-video cost the same. Based on the pricing fal lists, the cost of generating a 10-second video is as follows.
| Output resolution | Per-second price during discount | Cost of 10-second video | Cost of 10-second video at regular price |
|---|---|---|---|
| 720p | $0.0425 | $0.425 | $1.70 |
| 1080p (upscaled) | $0.0725 | $0.725 | $2.90 |
The regular prices are $0.17 per second for 720p and $0.29 for 1080p, and the amounts in the table are calculated by multiplying each per-second price by 10 seconds. During the discount period, you can try multiple ideas at a relatively low cost, but in actual production, failed footage may need to be regenerated, so the total cost of completing a video of the required quality should also be considered.
FLUX 3 Max's distinguishing feature is that it is said to sharply shorten waiting time for generation while maintaining and improving video quality. However, the speedup figures fal presented alone do not show how much time can actually be saved in a real production environment. If you are considering adoption, it is advisable to match the video length, resolution, and audio setting, compare reasoning_level at both low and high, and measure both the API's inference time and the total time until you receive the video.
If it can reliably produce footage of the required quality under the same conditions, FLUX 3 Max's speedup could change how trial and error is done in video production. Going forward, attention will focus on whether generation speeds close to the published figures can be achieved in real usage environments, and on how shorter generation time affects video quality and production costs.