The AI video models DV4 runs, and what each one does

DV4 turns a prompt into a finished MP4 with sound. No single model does that alone. One model plans the video. Another films the shots. Another composes the music. Code adds the words on screen. This page lists the models DV4 runs today, the ones that are coming, and the job each one does.

The lineup at a glance

What each model costs depends on your plan and the resolution you pick. See pricing for the current rates.

LTX 2.3: video with dialogue

LTX 2.3 comes in two versions, Fast and Pro. DV4 runs both through LTX's hosted API.

LTX makes the speech and the scene sound in the same pass as the picture. That means dialogue, ambience and effects. When a character speaks, the lines go into the shot prompt in quotation marks, and the model speaks them on screen.

LTX 2.3 Fast is the model on the Free plan. A free video is up to 15 seconds at 720p, with LTX's own sound. LTX 2.3 Fast is also on Lite, Plus, Pro and Team. LTX 2.3 Pro is on Plus and Team.

LTX 2.5: video up to 4K

LTX 2.5 also comes in Fast and Pro. It is the model for sharper, longer work, and it renders up to 4K. It is on the Pro and Team plans. LTX 2.5 Fast is the default model on both. LTX 2.5 Pro is there when you pick it.

LTX 2.5 can start a shot from a picture you upload, used as the shot's first frame.

MiniMax-H3 Optimized: another video engine

MiniMax-H3 Optimized is a second video engine. DV4 sends its shots to MiniMax's video API.

It is offered only outside the United States, the United Kingdom, the European Union and South Korea. If you use DV4 from one of those places, it does not appear in the model picker.

MiniMax-H3 makes video up to 2K. It does not make 4K. DV4 writes each MiniMax-H3 shot in that model's own prompt format, not the format it uses for LTX. Shots are fitted to your video's size and frame rate after they arrive, so they cut cleanly with the rest of the film.

Powered by MiniMax H3

Z-Image: pictures from a prompt

Z-Image makes still pictures. You type a prompt and get one PNG back, square, wide (16:9) or tall (9:16). Z-Image can draw English and Chinese words inside a picture. Every picture is checked by DV4's safety layer before you see it, words included.

ACE-Step: the music under every video

ACE-Step composes one instrumental track for the whole video. DV4 makes the music first, then cuts the shots to its beat. The music is mixed in at the end.

The track stays separate from the video model on purpose. Music played under speech during filming weakens lip sync. A separate track also plays unbroken across every shot. If you upload your own soundtrack on Lite, Plus, Pro or Team, it plays in place of the composed music.

The planning models

Before any shot is filmed, a language model reads your prompt and writes a plan. A paid plan holds the shot list, the music brief and the words for the screen. The planner is GPT-6.1 Sol.

Video models do not render text in DV4. Every shot prompt asks for footage with no readable words. Titles, callouts and end cards are added afterwards as code, so the spelling is exact. DeepSeek V4.1 designs where that text goes.

For motion graphics, DeepSeek V4.1 writes the whole film as code. No footage is generated. The scenes come with music.

How DV4 picks the engine for a shot

You pick a video model in the chat bar. Your plan decides which models appear there. A model your plan does not include shows a lock and the plan that unlocks it.

Once a video starts, every shot uses the model you picked. DV4 does not switch you to a different model halfway through. Each shot is made at a resolution your model supports on your plan. Shots are kept within the longest take that model makes.

Some choices follow the kind of shot:

• A shot where someone speaks gets its speech straight from the video model, in the scene's own sound.
• An uploaded soundtrack replaces the composed music; it is never used as a voice.
• A motion graphics scene skips the video model and renders as code.

DV4 also keeps characters consistent. Scenes with the same people that fit in one take become one shot with internal cuts. When the same people stay in the same place for longer than one take, the next shot starts on the last frame of the one before.

Why one product with several models

A single clip is rarely the thing you need. You need a video with a beginning and an end, voices that match the faces, music that fits the length, and words you can read.

Each model in DV4 does the part it is good at. The planner breaks your idea into shots. The video model films them with their own sound. ACE-Step scores the whole piece. Code places the text. Safety checks run on your request, the plan and the frames, on every plan.

You write one prompt. You get one finished video. To compare what the models make, see the showcase. To see which plan includes which model, see pricing.