This week fal released a set of numbers worth highlighting: using MiniMax's open-weight H3 as the base, H3 Max completes model inference for 5 seconds of 768p video in roughly 2.46 seconds — faster than the video itself plays. This isn't just a hardware upgrade; it's evidence that "model training + inference engineering + hardware" three-layer co-design is becoming a new paradigm. But we should also flag: 2.46 seconds is GPU inference time, not the end-to-end wait from a user click to seeing the video.
What this is
fal didn't train a video model from scratch. It built on top of MiniMax's open-weight H3. MiniMax H3 itself is a multimodal generative model: text, images, video, and audio can enter the same context, with the model understanding the relationships between materials.
fal did two things on this base. First, post-training (the foundation model already has full capabilities; new data and targets steer it toward a specific direction), adapting the model to fewer sampling steps. Video diffusion models start from noise and go through multiple denoising rounds to produce a result — the traditional approach uses dozens of steps; fal baked "lower inference cost" into the training objective, so the model maintains quality with fewer steps. Second, it built the Falcon inference engine in-house, covering quantization (using more resource-efficient numerical representations), parallelism, caching, and scheduling — squeezing the actual execution cost of each step. Three layers stacked, plus NVIDIA GB200 NVL72 clusters, gave us 2.46 seconds. fal calls this co-design: model training and inference optimization are no longer two independent stages, but from day one jointly search for the balance between quality, speed, and hardware efficiency.
Industry view
Mainstream take: this is a textbook example of "Chinese open-source base + overseas engineering team" collaboration. MiniMax opens the weights and hands over foundational capabilities; fal uses its engineering muscle to turn it into a commercial product. The division of labor is finer than before, and it represents the first complete commercialization of open-source models in video generation.
But three risks deserve attention. First, the numbers can mislead. fal's own API documentation defines 2.46 seconds as the GPU backend denoising time — excluding queuing, startup, weight loading, encoding, and transmission. Under high concurrency, real user wait times can far exceed this; be cautious about using this number in marketing. Second, the co-design bar is high — fal's own assessment: "hardware is buyable, inference engines require long-term system accumulation, post-training needs a team that can actually modify the model." Few teams have all three layers in place. That means the headline achievement can easily be mimicked as "stack more GPUs," while the real methodological moat is hard to transfer. Third, the commercial loop on open weights hasn't closed yet: MiniMax opens H3 and gets an engineering partner like fal, but whether the ecosystem can feed back into MiniMax itself remains unclear.
Impact on regular people
For enterprise IT: Video generation moving from "minutes" to "seconds" makes batch production scenarios — marketing, e-commerce, localization — viable. But buyers must distinguish "model inference time" from "end-to-end latency" and not get swept along by vendor demo numbers.
For individual careers: The content production ceiling for self-media, short video, and e-commerce operators is raised, and competition accelerates. Moats built on "speed" will be shorter-lived than before — the contest becomes judgment, not execution.
For consumer markets: Consumer users won't feel much in the short term — the few seconds of waiting inside video apps still exists. But within 6–12 months, AI video tools may shift from "trial toys" to everyday workflow.