Kandinsky 6.0 Video Is Open Source: What Ships and What It Takes to Run
On October 6, 2026, Kandinsky Lab open-sourced Kandinsky 6.0 Video, a family of diffusion models that generate video and audio together. The code, weights and Diffusers integration are released under the MIT licence, which makes it one of the most permissively licensed audio-video models available. This piece covers what is in the release, what it takes to run, and how it compares with the other open video models. Last checked: 2026-10-08.
What is in the release
- Two sizes. Kandinsky 6.0 Video Lite has 3 billion parameters, Pro has 29 billion. Each comes as a main checkpoint, a faster distilled checkpoint and a pretrain checkpoint.
- Video with sound. Both generate 5-second clips with synchronized 44 kHz audio, including lip-sync, from text (T2AV) or from text plus a reference image (TI2AV).
- Super-resolution add-on. A separate model, Kandinsky 6.0 Video Super-Resolution, upscales clips by 2x, 2.25x or 4x. That is how the release reaches Full HD (1920x1080); the base Pro example renders at 864x480 and 24 fps.
- Day-one tooling. Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo support, plus a Hugging Face Space demo of Pro Distill.
Licence
The GitHub code and every checkpoint on Hugging Face are MIT. There is no territorial restriction, revenue threshold or acceptable-use annex: you can self-host, fine-tune and ship commercial products on it. The one footnote is in the super-resolution repo, where the Diffusers export templates keep their original Apache-2.0 notices.
What it takes to run
Kandinsky 6.0 needs an NVIDIA GPU and Python 3.13 or 3.14. The repo ships presets from data-centre cards (H100, A100 80 GB, RTX PRO 6000) down to consumer cards (RTX 5090, 5080, 4090 and the 16 GB RTX 5060 Ti). Consumer cards run with block offloading, and the 16 GB preset uses the distilled Pro checkpoint with a quantised text encoder. Even 80 GB is not enough for Pro without CPU offload in vLLM-Omni. Plan for disk too: the Pro transformer weights alone are about 70 GB, against about 7.5 GB for Lite.
The bigger constraint is time. The repo's own benchmarks for one 5-second clip, measured after warm-up and excluding model loading:
| 5-second clip | H100 | RTX 4090 | RTX 5060 Ti |
|---|---|---|---|
| Lite, SD | 239 s | 437 s | 1,310 s |
| Pro, Full HD | 402 s | 1,247 s | 3,530 s |
That is close to seven minutes per clip on an H100 for Pro at Full HD, and about an hour on a 16 GB card. The distilled checkpoints cut this down, but this is a model for batch rendering and research, not for interactive previews on a laptop GPU.
How it compares with other open video models
| Kandinsky 6.0 Video | MiniMax H3 | Wan 2.2 | |
|---|---|---|---|
| Licence | MIT | MiniMax Community License; excludes the EU, UK, South Korea and US, and needs approval above $20M revenue | Apache-2.0 |
| Self-hosted output | 5 s; Full HD through the SR add-on | 768p; 2K needs MiniMax's hosted API | 480p or 720p |
| Audio | Generated, 44 kHz, with lip-sync | Generated | Not generated (the S2V model takes audio as input) |
| Smallest setup | 16 GB card with offload, slowly | Reference config uses 4 GPUs per variant | 5B model on a 24 GB RTX 4090 |
The short version: if you need open weights with sound and a licence your legal team will sign off on anywhere, Kandinsky 6.0 is now the obvious candidate. MiniMax H3 offers more resolution headroom, but its licence rules out self-hosting in four major markets. Wan 2.2 remains the lighter, well-supported choice when you do not need generated audio.
Is Kandinsky 6.0 on Imgo?
No. Kandinsky 6.0 is not integrated into Imgo, and we are not promising online generation for it. If you want hosted video with native audio today, Imgo runs Seedance 2.0, with clips of up to 15 seconds, image, video and audio references, and generated sound. The live model list shows everything you can generate right now.
Sources
Share
Related Posts
No image
Kling 4.0 Flash opened in early access on September 28 for a limited group — reported as Ultra yearly subscribers — with 3–20 second 720p clips. Here is how to get in, what it costs, how to write the 8,000-token prompts and keyframe plans the full model will accept, and what to generate meanwhile.
No image
Wan 3.0 accepts a document, a web link, images, videos and audio as inputs. This operator's guide covers what each input is for, how to structure a 30-second-take prompt, per-second budget examples at 480p to 1080p on both tiers, and what commonly goes wrong.
No image
Kling 4.0 Flash has been in early access since September 28, 2026; the full model and its API are pending, with an October 2026 target. Here is what each tier does, which headline specs are only announced, and what to use in the meantime.