Kandinsky 6.0 Video or a hosted API: a 5-second clip with sound

Kandinsky 6.0 Video is the better pick if you want open weights you can run yourself: it makes a 5-second clip with synchronized 44 kHz audio and lip-sync, under the MIT license. A hosted API is the better pick if you want no GPU to babysit. Sume does not list Kandinsky, but it lists two video ids that make sound natively: Gemini Omni Flash 1.1 (3 to 10 s) and MiniMax H3 Max (5 to 15 s).

Kandinsky Lab released the family on 2026-10-06 as a 29B Pro and a 3B Lite. The facts below come from the project repository and the technical report, both read 2026-10-11. Sume's side comes from Video Generation and Video Router.

What Kandinsky 6.0 Video says about itself

The repository lists text-to-audio-video and image-to-audio-video modes, and the report says a separate super-resolution model raises native SD output to 1920x1080 with scale factors of x2 and x4. The report gives a frame rate of 23 to 25 fps and describes the audio as dialogue, music and ambient sound that follow the picture. It calls the model competitive with leading audio-video generators, particularly in speech quality; that is the authors' claim, not something this post tested.

What Sume lists that makes sound natively

Sume's docs say the Gemini Omni Flash 1.1 id (gemini-omni-flash-1.1) accepts 3 to 10 seconds at 360p, 720p, 1080p and 4K, in 16:9 or 9:16, with native synced audio that is always on. MiniMax H3 Max (minimax-h3-max) runs 5 to 15 seconds at 480p, 768p or 1080p with native stereo audio. The prices are list x 1.25, taken from the rate card in Sume's pricing tests: Omni Flash bills $0.125 per second at 720p and $0.1875 per second at 1080p, and a default 5-second H3 Max request bills $0.50. The 5-second Omni figures are that rate multiplied by 5, not a separate quote.

How to choose

Pick on the constraint that binds first.

• Choose Kandinsky 6.0 when you need weights you control, want to tune settings yourself, or must keep prompts on your own machine. You also own the queue, the GPU and the install.
• Choose a Sume id when you want a job envelope (submit, poll or webhook, download) and a per-clip price with no hardware. Sume's video models report seed: false and reject a seed, so you cannot pin a rerun the way you can with your own pipeline.
• Choose by length: Kandinsky stops at 5 seconds, Omni Flash at 10 and H3 Max at 15. Longer needs a different id or a stitch.
• Choose by speech: Kandinsky generates dialogue from the prompt. If you already have a recorded voice-over, that is lip sync on a still, a separate Sume surface.

Find the matching ids in code

The video catalog is queryable, so you do not need to trust a blog table. This script prints every id that can generate audio and accepts a 5-second duration.

import os
import requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
r = requests.get("https://api.sume.com/v1/videos/models", headers=H, timeout=30)
r.raise_for_status()
for m in r.json()["data"]:
    ok_len = 5 in (m.get("supported_durations") or [])
    if m.get("generate_audio") and ok_len:
        print(m["id"], m["supported_resolutions"], m["pricing_skus"])