The Pulse
Kandinsky Releases Its 29B Video-and-Audio Model Under MIT
Kandinsky Lab open-sourced Kandinsky 6.0 on October 6, including a 29-billion-parameter model for generating video with synchronized audio. The model makes five-second clips from text or images, but the lab’s own tests show generation can t
AI.info Team · Written with AI ·
Kandinsky Lab released a 29-billion-parameter model that creates five-second video clips with synchronized sound, including lip-synced speech. The release, posted to GitHub on October 6, also includes a smaller 3-billion-parameter model, source code and checkpoints under the MIT license. Both models can work from a text prompt or an image paired with text; an additional super-resolution model can raise output to 1920-by-1080.
One model family, two sizes
The larger model, Kandinsky 6.0 Video Pro, and its smaller counterpart, Video Lite, generate video and audio together rather than requiring users to add a soundtrack in a separate step. In its paper, Team Kandinsky describes the output: “Both models generate 5-second video clips with synchronized 44 kHz audio, including lip-sync, in text-to-audio-video (T2AV) and image-to-audio-video (I2AV) modes; a built-in super-resolution model raises the output resolution to Full-HD (1920×1080).” Team Kandinsky et al., “Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation”
The technical report says the system connects a pretrained video stream with a newly trained audio stream in a dual-stream CrossDiT architecture. Bidirectional cross-attention links the streams to align meaning and timing. The authors describe training the audio stream first, then training both streams on paired audio-video data, followed by supervised fine-tuning, reinforcement-learning-based post-training and distillation.
Five seconds can take more than 20 minutes
The 29B parameter count describes the Pro model, not the full family; Lite is 3B. Kandinsky’s repository lists an NVIDIA GPU and Python 3.13 or 3.14 as requirements, and offers configuration for several GPU generations. It also provides ComfyUI extensions and says vLLM-omni support was added on October 6.
The repository’s own timings put a practical limit on what “available” means for local users. For the non-distilled Pro model on an RTX 4090, Kandinsky reports 936 seconds for a standard-definition five-second clip, 1,194 seconds for HD and 1,247 seconds for Full HD—roughly 16 to 21 minutes. Those figures follow warmup and exclude weight loading and MP4 encoding; they are the lab’s measurements, not an independent benchmark.
The comparison is not a clean sweep
Kandinsky’s paper reports that human evaluators preferred Video Pro to its predecessor, Kandinsky 5.0 Video Pro, and characterizes the new model as competitive with leading audio-video systems, particularly for speech. Its own comparisons also show trade-offs: the report says Veo 3.1 Fast leads on several audio and prompt-following measures, while Kandinsky performs better on some visual criteria. Against MiniMax H3, the authors report H3 ahead on most of the evaluated criteria.
The release gives researchers and developers code and model checkpoints to inspect, adapt and run under the stated license. But the speed figures make the distinction between access and convenient use plain: even the Pro model’s standard-definition clip takes about a quarter of an hour on the listed RTX 4090 setup, before loading and encoding time.