Video: MiniMax H3 - Timeline => New Axiris.ai service
16 août 2026 · Kris Wilhem
I have produced a complete sequence with the new MiniMax H3 video model, and above all with an in-house timeline for composing the shots. Nine shots, 46 seconds at 2520 × 1440, generated locally on a single machine.
What interested me was not releasing one more clip. It was answering a question nobody documents: what does it really cost? Every figure below is measured, not estimated.
A timeline, not a generator
Video generation services render a clip. They do not render an edit.
My timeline treats a film like a score: each shot carries its type of cut, its intent, its camera move, its soundscape and its start image. Each shot is launched on its own, judged, and redone if it does not suit me. The film is only assembled once I have validated everything.
A technical detail that changes everything in use: there is no disk cache between shots. Chaining clips is often done by carrying a state from one shot to the next: it is efficient, and it ties the shots to one another. Here each shot is independent. I redo shot 2 after validating shot 9, without recomputing anything.
What it costs
- 95 seconds to produce a shot, sound included
- 14 min 20 for the nine shots
- 1 h 13 for upscaling to 1440
- 1 h 28 in total, for 46 finished seconds
What surprised me most is in the report: creation costs a quarter of an hour, finishing costs seventy-three minutes. That is exactly why each shot is judged before anything is finished: upscaling is paid only once, at the end, on what has been validated.
Upscaling can be calculated
I measured its cost at three resolutions, same shot, same script, three passes in a row:
- 1024 → 1792 × 1024 · 1.92 s per frame
- 1080 → 1890 × 1080 · 2.21 s per frame
- 1440 → 2520 × 1440 · 3.84 s per frame
That is 1.06 seconds per megapixel produced, constant to ±1.8 % over a twofold change in surface area. No threshold, no fixed cost. Anyone can therefore estimate their own film before launching it, with a simple multiplication.
The hardware
RTX 6000 Pro Blackwell 96 GB, 128 GB of RAM. The times given assume a lightx2v distillation LoRA: four sampling steps instead of twenty. Publishing times without saying what makes them possible would make no sense.
The maximum-demand version
A second page presents three shots pushed to 96 frames per second. Social networks re-encode and cap the frame rate: the same video published elsewhere loses its fluidity. It is therefore hosted here, exactly as it comes out.
The timeline is still in development. It is inspired by the interface of LTX Director 2.0 (GPL-3.0) and was reimplemented from my own measurements, without reusing a line of its code, precisely to remain free of its licence.