What is Kandinsky 6.0 Video?
Kandinsky 6.0 Video was open-sourced on 6 October 2026 by Kandinsky Lab, part of Russia's Sber. It comes in two sizes, Lite with 3 billion parameters and Pro with 29 billion, each with standard, distilled and pretrain checkpoints. Both generate 5-second clips with synchronised 44kHz audio, including lip-synced speech, music and ambient sound, from text or from an image plus text; an update on 25 October added first-and-last-frame, audio-driven and video-driven modes.
The base clip is 480x864, and a separate super-resolution model takes it to 1920x1080. Code and weights are MIT-licensed. The developers say the audio module was trained on 40 million audio tracks and then jointly with the video module on 7 million paired clips.
Is Kandinsky 6.0 Video worth it?
Kandinsky 6.0 Video brings MIT-licensed video with synchronised speech and sound to ordinary graphics cards. Kandinsky Lab's own timings put a distilled Lite clip at 68 seconds on an RTX 4090 and 174 seconds on a 16GB RTX 5060 Ti, and a distilled Pro clip at 235 seconds on the 4090 or 88 seconds on an H100, all at 480x864 before upscaling.
Independent quality data is thin. Neither size appears on Artificial Analysis's text-to-video board with audio. The developers claim Pro beats LTX-2.5 on visual and speech quality and trails only MiniMax H3, while Digital Journal notes that results against LTX-2.5 and Veo 3.1 Fast varied by task and criterion. Reinforcement learning cut the word error rate of Pro's generated speech by 47%, according to the team.
Five seconds is a hard limit for now, and every clip is exactly that long, so longer scenes mean stitching. Full HD costs time too: a distilled Pro clip goes from 235 to 381 seconds on an RTX 4090 with super-resolution.
The maker matters for some buyers. Kandinsky Lab belongs to Sber, and Sberbank has been under full US blocking sanctions since April 2022. Downloading MIT weights from Hugging Face is not the same as paying Sber, but US companies should take legal advice before building a product on it.
Where it earns its place
A two-person language-learning start-up
Generate short talking scenes for vocabulary lessons
Run Lite Distill on a 16GB RTX 5060 Ti at about 174 seconds per 5-second clip with lip-synced speech; MIT terms mean no royalties as the app grows.
A social-media agency testing ad concepts
Produce many 5-second sound-on drafts quickly
Use fal's Pro endpoint at 8 steps for $0.24 a clip, so 500 drafts cost about $120, then re-render the chosen few at 50 steps with 1080p upscaling for $1.53 each.
An ML engineer at a Moscow media group
Self-host a domestic video model for production
Serve Pro on an H100 through vLLM-Omni or SGLang at about 88 seconds per distilled 480x864 clip, or 133 seconds with Full HD super-resolution.
Before you commit
- Setup
- Install uv and just, run just setup, download a checkpoint such as pro-distill and generate from a prompt; Diffusers and ComfyUI paths also exist.
- Integrations
- Diffusers, ComfyUI, vLLM-Omni, SGLang, FastVideo, a Hugging Face Space for Pro Distill, and fal's text-to-video and image-to-video endpoints.
- Data and privacy
- Self-hosted MIT weights keep footage local; training data beyond the 40 million audio tracks and 7 million paired clips is not described.
- Cost at scale
- fal: 1,000 Pro clips cost $1,350 at 50 steps or $240 at 8 steps; self-hosted, an RTX 4090 makes about 15 distilled Pro clips an hour at 480x864.
- Limits
- 5-second clips only; base 480x864 with optional upscale to 1920x1080; Pro needs block offloading on 24GB and 32GB cards; quality claims are self-reported.
What’s good
- MIT licence for code and weights
- Synchronised 44kHz speech, music and effects
- Lite runs on a 16GB RTX 5060 Ti
- Supported in Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo
What’s not
- Fixed 5-second clips
- No independent arena ranking yet
- Pro takes about 4 minutes per distilled clip on an RTX 4090
- Maker is part of a US-sanctioned bank
Best for
Developers and small studios who want commercially free, self-hosted short video with dialogue and sound, and can clear the Sber question with their lawyers.
Kandinsky 6.0 Video alternatives
Kandinsky 6.0 Video: common questions
- Is Kandinsky 6.0 Video worth it?
- Kandinsky 6.0 is a cheap way to get talking, scored video out of an MIT-licensed model, and the five-second ceiling is the price. We rate it 3.7 out of 5.
- How much does Kandinsky 6.0 Video cost?
- Free weights under MIT. fal charges $1.35 for a 5-second 480p Pro clip at 50 steps, $0.24 at 8 steps, and $1.53 with 1080p upscaling.
- What are the downsides of Kandinsky 6.0 Video?
- Fixed 5-second clips; no independent arena ranking yet; pro takes about 4 minutes per distilled clip on an RTX 4090; maker is part of a US-sanctioned bank.
- Who is Kandinsky 6.0 Video best for?
- Developers and small studios who want commercially free, self-hosted short video with dialogue and sound, and can clear the Sber question with their lawyers.
- How much does Kandinsky 6.0 Video cost for a team or in production?
- fal: 1,000 Pro clips cost $1,350 at 50 steps or $240 at 8 steps; self-hosted, an RTX 4090 makes about 15 distilled Pro clips an hour at 480x864.
- Is Kandinsky 6.0 Video safe for company data?
- Self-hosted MIT weights keep footage local; training data beyond the 40 million audio tracks and 7 million paired clips is not described.
- What are the best Kandinsky 6.0 Video alternatives?
- The closest alternatives we have reviewed are LTX-2.5, MAGI-2 Preview, MiniMax H3 (Hailuo 3.0) and HunyuanVideo 1.5.
Answers checked Nov 7, 2026.