Kandinsky 6.0 Pro beats LTX 2.5 and MiniMax H3 in a blind test of video models with audio

Choosing a video model based on polished demos is like choosing a laptop from the picture on the box. Every vendor showcases its best takes, while the failed attempts remain off-camera. Blind testing at least removes the brand effect: you judge the result, not the logo. One blogger conducted exactly this kind of experiment, and the outcome was unexpected for those accustomed to seeing open models as perpetual runners-up.

How the test was conducted

The creator selected three models that generate video with audio: LTX 2.5, MiniMax H3, and the newly released Kandinsky 6.0. Their names were hidden throughout 10 rounds, meaning he watched the clips without knowing which model had generated each one. Only afterward did he match the results to the model names.

Kandinsky 6.0 Pro won. According to the creator, it came out ahead in three areas: animation quality, the rendering of crowd scenes, and the stability of human anatomy. These are precisely the areas where video models usually fall apart: crowds turn into a blur, people's hands and faces drift, and motion looks jerky.

Why more than the victory matters

The key part of the news is the terms under which the model is available. Kandinsky 6.0 is distributed under the open MIT license. It is one of the most permissive licenses: the model can be used in commercial products without separate agreements. For a team building a video-generation service, that means fewer legal risks and no dependence on someone else's API or pricing.

The second detail is size. Enthusiasts have already optimized the Pro version down to 17 GB and released builds designed to run on standard hardware. This does not mean the model will run on any laptop, but the barrier to entry is considerably lower than with the heavyweight original weights. You can test the model on your own machine or an inexpensive rented GPU without budgeting for a cluster.

Limitations and what to do next

This is one test by one blogger across 10 rounds. The sample is small, the criteria are subjective, and the prompt set is unknown. Scene selection has a major impact on the outcome: a model that wins on crowds and anatomy may lose on in-frame text, long shots, or a particular style. The summary also says almost nothing about audio quality, even though the test specifically covered video with audio. It should therefore be treated as a signal that the model is worth investigating, not as a final verdict.

The practical takeaway is simple:

• If you are building a product with video generation, add Kandinsky 6.0 Pro to your comparison set alongside LTX 2.5 and MiniMax H3.
• Repeat the blind test on your own workloads: take 10–20 representative prompts, conceal the model names, and ask colleagues to rate the results.
• Evaluate audio, the stability of people in the frame, and generation time on your hardware separately.
• If the results are comparable, an open model under the MIT license offers an advantage in cost and control: you can deploy it on your own infrastructure.

The source is the blogger's YouTube video, where you can watch the clips and judge the differences yourself. If your own eyes and prompts confirm his conclusion, it will be a strong argument for an open video stack.