On October 6, 2026, Sber enabled video generation with sound in GigaChat: the neural network draws the visuals and adds its own audio—including voice, music, and ambient noise. For GigaChat users, it's free, with no credit card required and no strings attached. The clips are up to five seconds long and can be rendered in up to Full HD quality.
For short-form video creators, this means there's now an AI video tool that works from Russia, perfect for inserts, intros, and draft UGC scenes. Let's take a look at what it can do, what it can't, and how to fit these five-second snippets into a real video.
What Exactly Was Launched?
The generation is powered by Sber's new model, Kandinsky 6.0 Video. Its main difference from previous versions is that the image and sound are created together by a single model, rather than being stitched together afterward. This means a character's lips move in sync with their speech, footsteps match their stride, and the music shifts with the scene.
You can prompt the scene in two ways:
- Describe it with text—what's happening, who's speaking, what sounds are needed.
- Upload a starting frame (your own photo or an image) and separately describe what should happen and how it should sound.
The sound is optional; the model can also create silent videos.
Features: What Sber Promises
| Parameter | Value |
|---|---|
| Clip duration with sound | up to 5 seconds |
| Quality | SD, HD, or Full HD (up to 1920×1080) |
| Audio frequency | 44 kHz |
| Frames per second | ~24 fps |
| What the audio includes | lip-synced voice, music, ambient noise, effects |
| Price for GigaChat users | free |
| Generation time | ~8 minutes (according to ppc.world) |
Five seconds is the current ceiling. Sber says it plans to increase this but hasn't given a timeline.
How Does It Stack Up Against Competitors?
It's important to note where these numbers are coming from. All the comparisons below are Sber's claims; there are no independent benchmarks yet:
- The new model outperforms the previous one, Kandinsky 5.0, in about 71% of cases across all criteria.
- In visual quality and motion accuracy, it surpasses the open-source model LTX 2.5 and the closed-source Veo 3.1 Fast.
- The team's scientific paper is more cautious, stating that Kandinsky 6.0 Video is ahead of LTX 2.5 on most VABench test metrics and "remains competitive" with proprietary systems, especially in voice quality.
Regarding training data, Sber reported "over 20 million videos with sound." The paper provides a more detailed breakdown: 50 million images, 20 million video scenes, 40 million audio tracks, and 7 million fragments where video and sound are synchronized.
The Model Was Also Released to Developers for Free
In separate news, Sber has open-sourced the model under an MIT license. It's the most permissive of the common open-source licenses, allowing for use in commercial products.
"We are providing the Kandinsky 6.0 Video model to developers for free and without restrictions," said Anton Frolov, Senior Vice President and Head of Generative AI Development at Sberbank.
What was released:
| Item | Details |
|---|---|
| Model versions | Lite — 3 billion parameters, Pro — ~30 billion parameters |
| Full HD upscaling | a separate 1-billion-parameter model, ×2 and ×4 |
| Where to find it | Hugging Face and GitHub |
| Integrations | Diffusers, FAL, FastVideo, ComfyUI |
| Local execution requirements | from 16 GB of VRAM (according to Habr); the model card allows offloading parts to RAM to fit on smaller cards |
A useful technical detail: the model generates video at a small resolution—864×480. The Full HD quality comes from upscaling with a separate model. So, "Full HD" here is an enlarged image, not a native 1080p render. For short inserts, the difference is barely noticeable, but when stretched to full screen, the softness becomes apparent.
How to Use This for Short-Form Video
Five seconds isn't a full video—it's a detail. Here's where those details can work wonders:
Intros and Outros
A five-second scene with the right tone and sound makes a perfectly good intro for a series. Create it once and use it in every episode.
B-Roll for Voiceovers
When you're narrating something but have nothing to show, a five-second clip can fill the visual gap. In this case, just lower the model's audio and keep your own voiceover.
Concept Drafts for Clients
If you're making UGC and need to get a concept approved, showing a five-second scene is easier than explaining it. The actual shoot will go faster because everyone has already seen what you meant.
Bringing Photos to Life
The "start frame + description" mode is perfect for animating a still image. It works even better with stock photos, screenshots, and product shots than generating from scratch.
Editing is simple: cut 4–6 different five-second scenes, stitch them into a video, and lay a single audio track over the top. This way, the viewer stops noticing the length of each individual clip, and the music helps blend any stylistic differences between the scenes.
Don't Forget the AI Label
TikTok, YouTube, and Instagram require you to label realistic AI-generated content. The label is applied on the platform during upload, not in GigaChat. If your insert could pass for a real recording, label it: platforms will throttle your reach if you miss it, and in serious cases, they might even demonetize your content. Abstract graphics and clearly illustrated scenes are exempt from this requirement.
What's Still Unclear
Frankly, here's what Sber didn't explain:
- How many daily generations are allowed with free access. The word "free" is in the press release; the limit is not. We'll have to test this out in practice.
- Whether paid plans will be introduced later. Nothing has been said on this.
- Vertical format. The model cards and examples show horizontal resolutions (864×480 and 1920×1080). There has been no mention of 9:16, the format needed for Shorts, Reels, and Clips. For now, expect to have to crop a horizontal frame.
- The terms of use for clips made within GigaChat. The MIT license covers the model you download. What you can do with video created inside the GigaChat service is determined by the service's rules, and no separate clarification was issued with this launch.
- When the 5-second limit will be increased. The plans have been announced, but the dates have not.
If any of these points are critical for you, don't build your workflow around it until Sber provides clear answers.
FAQ
What free AI video generator works from Russia?
GigaChat with the Kandinsky 6.0 Video model. It's a Russian service that's easy to access, and video generation is free. We reviewed other alternatives and their availability in Sora is Shutting Down: Why and What to Replace It With — AI Video Alternatives for Shorts, Reels, and UGC.
How do you make a video with sound in GigaChat?
Describe the scene with text and specify what should be heard: who is speaking and what they're saying, what music to play, what noises to include. Alternatively, upload a starting frame and add a description of the action and audio. Choose the quality: SD, HD, or Full HD. A generation takes about eight minutes.
How long is a GigaChat video?
Up to five seconds with sound. They promise to increase the limit, but with no timeline. For a finished video, you'll need to stitch several five-second scenes together.
Can you post this video on TikTok, Shorts, and Reels?
Technically, yes. Two caveats: the frame is horizontal and will need to be cropped for the vertical format, and a realistic AI scene must be labeled as such when you upload it to the platform. You can find tools for editing and finishing your video in our roundup of AI Video Editing Software: 15 Best Editors of 2026, Free and Paid.
Where can I download Kandinsky 6.0 Video and what do I need to run it?
The weights and code are on Hugging Face and GitHub under an MIT license, with integrations for Diffusers, ComfyUI, FAL, and FastVideo. To run it locally, you need at least 16 GB of VRAM, according to Habr; the 3-billion-parameter Lite version requires less than the Pro version. If you don't have the hardware, using GigaChat is the easiest option.
Do I need subtitles if the model generates the audio?
Yes. The model's voice clips are short, and short-form video is most often watched without sound. How to make subtitles for free: Free Video Subtitles: X Now Corrects and Translates Them with Grok.
Free tools like this lower the barrier to entry: a scene that used to require a paid shoot can now be generated in eight minutes. On Prime Oracles, brand briefs are available for anyone to see: in the Content Rewards section, you can view the terms and the pay-per-view rate, and pick up a task by clicking the Accept Task button.
Sources: Sber press release (CNews), ixbt, Habr, ppc.world, models on Hugging Face, team's scientific paper.