fal Co-Designs Post-Training and GB200 Inference for H3 Max, Cutting Five-Second Video Generation to About Three Seconds

fal Research has released H3 Max, building on MiniMax H3’s open weights with post-training on new data targeting prompt adherence and visual quality, while its inference team simultaneously optimized the execution stack. The base H3 model is an approximately 33B-parameter multimodal video system capable of generating visuals and stereo audio in a single pass. H3 Max’s public API currently supports text-to-video and image-to-video generation, 480p and 768p resolutions, durations from five to fifteen seconds, and multiple aspect ratios.

The technical focus of this update is not simply reducing the number of denoising steps. fal says the team co-developed the checkpoint and core inference path on NVIDIA GB200 NVL72 systems, conducting preference tests on optimizations such as reduced precision, approximations of expensive operations, and the removal of sampling steps. It retained only changes that did not lower quality rankings. In fal’s official tests, a five-second video completed in about three seconds, with throughput approximately 35 times that of MiniMax’s official H3 endpoint; the company also claims it is, on average, 15 times faster than models of comparable quality. This brings video-model latency close to a range suitable for interactive editors, preview tools, and iterative agent-based revision loops.

For quality evaluation, fal conducted pairwise human-preference comparisons across twelve models, separately assessing overall preference, prompt understanding, and aesthetics. It aggregated the results using Bayesian Elo ratings with 95% confidence intervals, and its internal results ranked H3 Max first in all three categories. However, the live Artificial Analysis leaderboard for text-to-video models with audio places it third, with an Elo rating of 1,235 and a 95% interval of plus or minus 10 points; its interval overlaps those of first-place Wan 3.0 and second-place Gemini Omni Flash. This indicates that H3 Max belongs in the leading tier, but does not establish a consistent and statistically significant lead.

Deployers should also distinguish between an “open base model” and an “open derivative model.” MiniMax H3’s weights are downloadable, but fal’s announcement currently provides only the H3 Max Playground, Agent, and API. It has not released the derivative weights, post-training data, core code, or reproducible latency benchmarks. The API’s `timings.inference` field also reports DiT time only on certain routes and cannot be treated as a direct measure of queuing, prompt expansion, transmission, or full end-to-end latency. The next things to watch are whether fal releases the weights and inference recipe, and how costs scale across different video durations, batch sizes, and hardware configurations.