The cloud is no longer the only place to try this
There is a particular moment when a technology stops being a demo and starts becoming a tool. For local AI video, that moment is not about producing a perfect feature film. It is about asking a small machine for a short scene, pressing start, and receiving both moving images and a spoken soundtrack without sending the generation to a commercial service.
That is what we tested with MiniMax H3. The model is an open-source multimodal system from MiniMax that generates video and audio together. We installed it in an isolated ComfyUI environment on our local AI server, using an NVIDIA RTX 5060 Ti with 16 GB of VRAM. Nothing was rendered in a cloud API, and the existing image-generation service on the server was left untouched.
For European creators, that distinction matters. A locally generated draft does not require uploading every prompt, reference image or unfinished idea to a US-based platform. It does not remove all legal responsibilities, but it gives a filmmaker, educator or small business a much clearer place to start when privacy and control matter.
Our test produced video, dialogue and stereo sound
We ran two renders. The first was a control run at 864 × 480. The more ambitious test used a resolution of 1344 × 768 and produced a 5.17-second clip at 24 frames per second. We placed an Italian dialogue directly in the prompt.
The result contained a video track and stereo audio. Technical inspection showed a 32 kHz, two-channel audio stream, matching the output format described in the official MiniMax H3 repository. The generation completed without an out-of-memory error. VRAM use was approximately 12.2 to 14.1 GB, which is remarkably close to the practical limit of a 16 GB card.
The output is short, but that is exactly why the result is interesting. A five-second clip is enough to test a character, a camera move, a line of dialogue or a visual idea. For a small European studio, an independent creator or a school experimenting with AI media, the ability to iterate locally can be more valuable than a polished 30-second clip that requires a monthly subscription and a queue.
The face is convincing enough to keep watching
The visual result surprised us in a positive way. The character keeps a stable identity across the clip, the face does not collapse into obvious frame-to-frame distortions, and the mouth moves in a way that reads as speech. It is not indistinguishable from filmed footage, but it is far beyond the old local-video experience of flickering faces and broken hands.
This is where the technology becomes useful rather than merely impressive. A creator can use the output as a storyboard, a pitch visual, a teaching example or a first draft. A marketing team can test whether a concept works before booking a shoot. A teacher can demonstrate how prompts affect camera movement and dialogue without opening a student account on a remote platform.
We should be honest about the limitation: the audio is weaker than the image. The dialogue is present and the video includes synchronized sound, but the voice does not yet have the polish of a finished voice-over. We also did not verify cloning a specific reference voice. The workflow used in this test generated a voice from the prompt; it did not reproduce a supplied recording of a particular person.
Thirty-four minutes for five seconds
The main cost is not only hardware. It is time. The 1344 × 768 render took approximately 34 minutes for 5.17 seconds of video. That is considerably slower than the 11.4-minute figure reported in the post that inspired our test. A different workflow, cache behaviour or more aggressively optimised configuration could explain the gap, but we are reporting the result we actually measured.
The smaller control render took about eight minutes. Offloading and streaming make the larger configuration possible, but they do not make it fast. This is not a production pipeline for generating hundreds of clips overnight on one consumer GPU. It is a local creative instrument: slow enough to require patience, inexpensive enough to invite experimentation.
What European users should know
MiniMax H3 is available to European developers as an open-source project, and the local ComfyUI route does not depend on a country-specific web subscription. The model documentation lists stable dialogue support for several languages, including Italian, English, German, French, Spanish and Portuguese. Czech is not listed among the languages with stable support, so Czech voice quality should be treated as unverified rather than assumed.
Running the base generation locally can also reduce data exposure under the GDPR. Prompts, intermediate files and generated clips can remain on infrastructure controlled by the user. That is not a blanket compliance certificate: teams still need to handle personal data, copyright and consent correctly, especially when they use faces or reference audio. The practical difference is that the first processing step can happen inside the organisation instead of automatically leaving for a third-party cloud.
The AI Arena now records this run as a video-generation result rather than pretending that H3 is an ordinary text LLM. That separation is important. Text models are usually compared by tokens per second and coding quality. H3 needs different measurements: render time, resolution, frame rate, VRAM, image stability and audio quality.
Our verdict: the local future is already useful
MiniMax H3 is not perfect. The voice needs work, reference-voice cloning was not demonstrated, and the render time makes serious production impractical on this hardware. But the central result is hard to dismiss: a single 16 GB consumer GPU created a moving scene with dialogue and stereo audio locally.
That changes the conversation. Local AI video no longer has to mean a technical curiosity reserved for people with a data centre. It can be a patient, slightly rough, surprisingly capable tool on a workstation that many creators could realistically own. The image quality is already good enough to explore ideas. The next improvements do not need to make the technology magical; better audio and faster rendering would be enough to make it considerably more useful.
Can MiniMax H3 run locally on a 16 GB GPU?
In our ComfyUI setup, yes. The 1344 × 768 test completed on an RTX 5060 Ti with approximately 12.2–14.1 GB of VRAM used. Results will depend on the workflow, quantisation and available system memory.
Does the local workflow clone a person’s voice?
That was not demonstrated in our test. We used prompt-generated speech rather than a reference recording. Voice cloning would require a separate Ref2VA workflow and additional testing.
Is local H3 practical for professional video production?
Not yet on this hardware for high-volume work. Our five-second 1344 × 768 clip took approximately 34 minutes to render. It is already practical for prototypes, storyboards, experiments and short concept clips.
Sources: MiniMax H3 official repository; our own local ComfyUI test on an RTX 5060 Ti 16 GB.
