AI model
Video generation model family supporting text-to-video, image-to-video and reference-to-video, with an open-weight base described as generating at 768p.
The September 14th VC-Attention paper reports 1.46x on an H200 and 1.59x on a B200 against BF16 FlashAttention-4, measured at the attention call. How much of that reaches a finished clip depends on how much of your step is attention.
Reality
The speed figure comes from a single dev.to write-up and describes generation, not delivery. The design decision underneath it is that frames and stereo audio leave in one pass, so the smallest thing you can retry is the whole clip.
Reality
NVIDIA Research rebuilt the serving stack for someone else's open-weight video model and got a five-second clip out in 1.653 seconds, but the default profile is lossy and the dense comparison run is the honest baseline.
Reality
Atlas arrived Sept. 1 in early access with no price, no named partner and no paper, and the reconstruction win it publishes runs against a baseline whose own authors asked evaluators to wait. That makes it a procurement question.
Reality
The vLLM benchmark shows a finished MP4 arriving before its own playback would end, which is a different property from showing frames as they are made, and MiniMax's community licence still requires separate permission for US, EU, UK and Korean use.
Reality
The five-second clip in three seconds is fal's own figure, measured on fal's hardware against an endpoint it does not control, and the throughput multiple only transfers if you know the batch size and GPU count behind it.
Reality
fal's H3 Max clears real-time playback by two thirds at the fast end of its range, which turns continuous video into a metered compute bill. The 12.5-cent promotional clip price is what the arithmetic rests on.
Reality
Enterprise inference sales grew about eightfold to $73.9 million, and the reported loss came in $114 million inside consensus while the adjusted loss more than doubled. Both are true.
Reality
The integration index documents a real prototype-to-serving path with the weights unchanged. The step that reads a user's inputs is still an API call to MiniMax.
Reality
Causal 3D autoencoders quantise clip length to latent-frame boundaries, so only a short arithmetic progression of durations is reachable. Snap the number before you show, price or store it.
Reality
MiniMax's open-weight H3 base generates at 768p, and the 2K stage is not in the release. Pipelines validated on text-to-video demos will meet ceilings the demos never showed.
Reality