fal and Black Forest Labs launch FLUX 3 Max video model: tied #1 in image-to-video quality, a 10-second clip in about 12 seconds, roughly 22x faster than Seedance 2.5

Background and the problem it targets

Video generation has struggled to deliver quality and speed together: high-quality models such as Seedance 2.5 and Wan 3.0 often need minutes for a 10-second clip, so creators draft on fast models and finish on slow ones. On Oct 19 fal released FLUX 3 Max to show that targeted post-training and inference optimization on an open foundation model can reach top-tier quality at close to real-time speed. Black Forest Labs CEO Robin Rombach said in the press release that FLUX 3 was built as a foundation for others to build on.

Architecture and how it works

FLUX 3 Max is not trained from scratch. fal Research took Black Forest Labs' FLUX 3 and applied post-training, reinforcement learning and inference-system optimization, aiming at visual fidelity, motion coherence and prompt adherence. Text-to-video and image-to-video share one post-trained model, with a keyframes mode as well. The API supports 5-20 second clips, 720p or 1080p, several aspect ratios, audio generation on by default, and a reasoning_level setting (low/high) for how much effort goes into interpreting the prompt. fal says it will keep post-training the model.

Benchmarks and measured results

The third-party figure comes from Artificial Analysis's image-to-video arena (AA-Video-I2V v1.1, preliminary): FLUX 3 Max at 1,139 Elo (4,868 votes per the press release), level with Seedance 2.5, with MiniMax H3 on top at 1,141; the "tied #1" claim rests on overlapping confidence intervals. Speed numbers are fal's own (vendor-reported): 12.45s for a 10-second 720p clip versus 47.37s for Gemini Omni Flash 1.1 and 277.80s for Seedance 2.5. The press release says 10.95s, which differs from the product page; this article uses the product page. Against the original FLUX 3, Elo rose 50 points, from 1,089 to 1,139.

Getting started

1. Try it on the fal product page or model pages; API model IDs are blackforestlabs/flux-3-max/text-to-video and blackforestlabs/flux-3-max/image-to-video.
2. Call it through official SDKs such as @fal-ai/client, billed per output second: until Nov 7, $0.0425/s at 720p and $0.0725/s at 1080p, then $0.17/s and $0.29/s.
3. Set reasoning_level to low when iterating quickly on shots, and switch to fal's H3 Max on the same API to compare.