fal opens FLUX 3 Max for video generation from text, images and keyframes
fal has launched FLUX 3 Max, a video model adapted through post-training with Black Forest Labs. Developers can generate video from text, images or keyframes through a unified interface. The work combines additional model training with changes to inference and serving systems. Access is through fal’s hosted service, and the research team is continuing post-training after this first release.
Artificial Intelligence··Morning
FLUX 3 Max follows joint post-training
fal, a company providing hosted AI inference, launched the FLUX 3 Max video model on October 19 after work with model developer Black Forest Labs.[1], [2]
The starting point was Black Forest Labs’ FLUX 3 foundation model. fal Research then conducted post-training with the original developer, adapting that foundation for the released service. The collaboration links the development of the underlying model with further training and the engineering needed to run it for users. The model is now offered through fal’s inference platform, where developers send requests to a hosted system rather than running a downloaded model on their own machines.[1]
Three input routes share one interface
The application programming interface, or API, supports text-to-video, image-to-video and keyframes-to-video generation. A developer can therefore start from written instructions, an image or selected frames through the same access interface. These are the video-generation workflows named in the release. They describe how material enters the model, while the hosted inference service handles the request and returns its output. The October 19 release provides access to this first version through fal. It does not include downloadable weights, the numerical parameters needed to run the model independently, or a package for installation on a local machine.[1]
Training and serving changes continue after launch
fal describes the adaptation as a combination of additional training, reinforcement learning, inference optimization and systems engineering. Reinforcement learning forms part of the training work; inference optimization concerns execution of the trained model. Serving-system changes address the path through which production requests reach that model. The company identifies visual fidelity, coherent motion and adherence to instructions as goals of this work. Its launch also cites human-preference evaluations, while its speed comparisons use vendor measurements. The published latency values differ across fal’s materials and do not supply an independently established speed advantage. fal calls this the first iteration and says its research team is continuing post-training. For human-preference comparisons, the company cites Artificial Analysis evaluations of model outputs. It also notes overlapping confidence intervals among highly ranked models, limiting categorical claims of superiority. Those quality evaluations and the vendor’s own speed measurements are separate bases for comparison.[1]
