Blog · 2026-11-20
LTX 2.5 vs MiniMax H3: LTX Was 4 to 9 Times Faster, H3 Stuck Closer to Every Prompt
Model pages: LTX 2.5 Fast API · MiniMax H3 API — pricing, specs and the playground
The short answer: pick LTX 2.5 Fast for speed, many takes, your own voice-over or a fixed first frame; pick MiniMax H3 when the finished shot must match the prompt. On three identical 6-second prompts, LTX 2.5 Fast finished in 27 to 64 seconds and H3 in 4.1 to 4.8 minutes, yet H3 followed the staging, the shot order and the product more closely every time. A clip cost $0.486 on LTX at 720p and $0.582 on H3 at 768P.
LTX 2.5 is the newest video model from Lightricks, offered as a Fast tier and a Pro tier, with sound generated together with the picture. MiniMax H3, which many people call Hailuo H3, is MiniMax's multimodal video model: one request can carry a prompt plus up to nine reference images, three reference videos and three reference audio clips, and it returns 5 to 15 seconds of video with sound at 768P or native 2K. The two lean different ways. LTX 2.5 is built for speed and can build a clip around an audio file or a first frame you supply; H3 is built to follow a detailed description and a stack of references.
A word on who we are: we run apimodels.app, a platform where one API key calls LTX 2.5, MiniMax H3 and other image, video, audio and language models. Every clip in this article was generated on that platform on 20 November 2026. Apart from joining each pair into one player and labelling the model in the corner, nothing was edited. Spoken lines were checked by transcribing the audio.
How we tested: three everyday jobs with word-for-word identical prompts, namely a two-person dialogue scene, a three-shot action sequence and a vertical product clip. LTX 2.5 Fast ran at 720p and MiniMax H3 at 768P, its lower tier, both for 6 seconds with sound, once each, and we kept whatever came back. H3 delivered 6.6 seconds at 1344 × 768 per clip; LTX delivered exactly 6 seconds at 1280 × 720. Times run from creating the job to the finished file on our platform, queue included. We skipped image-to-video: on our API, H3 takes reference images rather than a first frame, so the two would not be doing the same job.
Results at a glance
- Dialogue: both models spoke the two lines word for word; only MiniMax H3 put the cook behind the counter as written. LTX 2.5 Fast took 41 seconds, H3 4.1 minutes.
- Three-shot action: H3 delivered the three shots in the order we wrote, but the puddle became a wet patch with no splash; LTX 2.5 Fast made the splash but opened on a wide shot instead of the tracking shot. 27 seconds against 4.2 minutes.
- Vertical product clip: H3 looked like a finished ad, with a dark stone, clear ice cubes and dark coffee, though its label is made-up text; LTX 2.5 Fast gave a pale background, lumpy ice and coffee the colour of tea. 64 seconds against 4.8 minutes.
- Price per 6-second clip: $0.486 on LTX 2.5 Fast at 720p, $0.582 on MiniMax H3 at 768P.
- What only one of them does: LTX 2.5 Fast can use your voice recording as the soundtrack and start from your first frame; H3 takes reference videos and audio, runs to 15 seconds and renders native 2K.
Put simply, MiniMax H3 won on how closely each clip followed the prompt and LTX 2.5 Fast won on time, by a wide margin. Neither was flawless: H3 wrote made-up text on the lanterns and the bottle label, and LTX bent the bottle and turned the ice into lumps.
Which one fits your job
- You generate many takes, previews or variations, or people wait on screen for the result: LTX 2.5 Fast. About a minute per clip in our test, against four to five minutes.
- The shot must match a detailed brief, such as who stands where, the order of shots or how a product looks: MiniMax H3. It was closer to the prompt in all three tests.
- You have a voice-over and need someone on screen to say it: LTX 2.5 Fast audio-to-video. Your recording becomes the soundtrack.
- You must start from a specific picture, such as a product photo: LTX 2.5, which takes a first frame. On our API, H3 takes reference images, which guide the look but do not fix the opening frame.
- You want reference videos or audio to hold a character, a camera move or a voice steady: MiniMax H3.
- You need 21:9, 4:3 or 3:4: MiniMax H3. LTX 2.5 offers 16:9, 9:16, 1:1, 3:2 and 2:3.
- You need 48 or 50 fps, or one clip longer than 15 seconds: LTX 2.5 Fast, up to 20 seconds.
How LTX 2.5 and MiniMax H3 differ on paper
- Resolution: LTX 2.5 Fast 720p, 1080p and 1440p; LTX 2.5 Pro 720p and 1080p; MiniMax H3 768P and native 2K, with 2K the default if you do not set one.
- Length: LTX 2.5 Fast 6 to 20 seconds in even steps, but only 6, 8 or 10 at 1440p or 48 and 50 fps; LTX 2.5 Pro 6, 8 or 10; MiniMax H3 any whole number from 5 to 15.
- Aspect ratio: LTX 2.5 16:9, 9:16, 1:1, 3:2 and 2:3 for text-to-video; MiniMax H3 adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.
- Frame rate: LTX 2.5 offers 24, 25, 48 or 50 fps; MiniMax H3 delivered 24 fps in our tests.
- What you can feed it: LTX 2.5 takes a prompt and one first-frame image, plus an audio file on Fast; MiniMax H3 takes a prompt with up to nine reference images, three reference videos and three reference audio clips.
- Sound: both generate sound with the picture; only LTX 2.5 Fast can use your audio file as the finished soundtrack.
- Open weights: both have published model weights, each under its own licence. LTX says LTX 2.5 is free for organisations under $10 million in annual revenue. MiniMax's public H3 weights are the base model at 768p, without the 2K stage, under the MiniMax H3 Community License, and the model card sends users in the US, EU, UK and South Korea to an application form. This article tests the hosted API only.
Dialogue: both said the lines, only MiniMax H3 kept the counter
In the ramen shop prompt, both models spoke the two lines word for word, and only MiniMax H3 staged the scene the way we wrote it. The prompt asked for a static medium shot of a tiny ramen shop at night: a cook in his fifties wipes the counter, looks up at a young woman sitting across from him and says "Last bowl of the night, on the house", and she laughs and replies "You say that every night", with simmering broth and rain on the window.

MiniMax H3 put the cook behind a long counter with a towel under his hand and a steaming pot in front of him, and the woman eating noodles on the other side, under paper lanterns with rain on the glass. It looks like a still from a Japanese TV drama. The lanterns carry characters that look Japanese but are not real words. LTX 2.5 Fast sat the two of them at the same small table; the cook does wipe it and the lines land on time, but the counter is gone. LTX 2.5 Fast took 41 seconds and MiniMax H3 took 4.1 minutes.
If the setting must look like a real place, do not ask either model for signs, lanterns or labels with writing on them; add the text afterwards. With LTX 2.5, state positions plainly and repeat the ones that matter, for example "standing behind the counter, across from her".
Three shots in six seconds: MiniMax H3 kept the order, LTX 2.5 made the splash
In the cyclist prompt, MiniMax H3 followed the shot list in order, while LTX 2.5 Fast delivered the most dramatic moment but shuffled the shots. We asked for a cyclist in a yellow jacket on a coastal cliff road at sunrise: a side tracking shot, then a low close-up of the front wheel splashing through a puddle, then a wide high shot of the road curving along the sea, with wind, a clicking chain and waves on the soundtrack.

MiniMax H3 opens on the side tracking shot, cuts to the front wheel low on the road and ends on a high view of the cliff road with the rider small in the frame, which is exactly the order we wrote. The rider, the yellow gloves and the bike stay the same across the cuts. The puddle, though, became a wet patch that reflects the wheel, with no splash. LTX 2.5 Fast opens on a wide shot into the sun, then the side view, then a big splash through the puddle, and passes through a muddled frame where the rider seems to ride along the cliff edge before the high shot. For sound, LTX gave a whoosh and the splash, H3 a low hum and a whoosh; neither track has a clear chain click or waves. LTX 2.5 Fast took 27 seconds and MiniMax H3 took 4.2 minutes.
Vertical product clip: MiniMax H3 looked like an ad, LTX 2.5 drifted
For the 9:16 product clip, MiniMax H3 came close to a usable ad, while LTX 2.5 Fast drifted on the background, the ice and the colour of the coffee. We asked for a glass bottle of cold brew coffee on a wet black stone with cold mist around it; a hand twists off the cap with a crisp hiss and pours the coffee over ice into a tall glass while the camera slowly pushes in.

MiniMax H3 shows the dark stone, the mist and a hand gripping the cap, then cuts to a tall glass of clear square ice cubes as dark coffee pours in. Two things are off. The label looks like a brand but is made-up text, and the twist itself is skipped: the hand grips the cap and the next shot is already pouring. LTX 2.5 Fast set the scene in pale fog instead of against a dark backdrop, the ice is white lumps, the coffee looks like tea, and the bottle changes shape between shots. Both soundtracks have the pour and a fizz, and H3 also has a pop near the start. LTX 2.5 Fast took 64 seconds and MiniMax H3 took 4.8 minutes.
For a real product, do not count on either model to write your label. MiniMax H3 accepts up to nine reference images, so you can send photos of the actual bottle; LTX 2.5 can start from a photo of it as the first frame. We did not test how faithfully either route reproduces a label.
Only LTX 2.5 can do this: your voice-over and your first frame
LTX 2.5 Fast has an audio-to-video mode that MiniMax H3 does not offer in this form: you send an audio file and, optionally, a picture, and the clip is built around your recording, which stays as the soundtrack. We sent a 4.8-second voice line and an AI-generated portrait of a podcast host, with a one-line prompt about her talking warmly to camera.
The clip ran 4.8 seconds, the same as the audio, came back in 86 seconds and cost $0.39, because without a set length the clip is billed by the length of your audio. The mouth follows the words. MiniMax H3 accepts reference audio too, but as a guide for the voice and sound of a new soundtrack; we did not test whether it reproduces a recording word for word.
LTX 2.5 also takes a first frame: send a product photo and the clip starts from that exact picture, which our LTX 2.5 vs Seedance 2.0 test shows with a coffee mug. On our API, MiniMax H3 takes reference images instead, which steer the look of a character or product without fixing the opening frame. If you need a first and a last frame in the H3 family, MiniMax H3 Lite has both, as a draft tier.
Only MiniMax H3 can do this: steer a shot with reference videos, audio and images
MiniMax H3 takes far richer input than LTX 2.5. One request can carry up to nine reference images, three reference videos and three reference audio clips, and the prompt says what each one should supply: a face, a camera move, a rhythm or a voice. LTX 2.5 takes one first-frame image and, on Fast, one audio file, and no other visual reference. H3 also runs to 15 seconds in any whole number of seconds, renders native 2K and offers 21:9, 4:3 and 3:4. We did not test reference inputs for this article; the MiniMax H3 docs show the request format, and How to write MiniMax H3 prompts covers how to tell H3 what each reference is for.
Reference material on MiniMax H3 is priced on top of the output seconds: the first five reference images are free and each further one costs $0.054, a reference video is billed for its own length ($0.135 per second at 2K), and reference audio is free. Every tier and resolution is laid out in MiniMax H3 API pricing, compared.
For better-looking shots: LTX 2.5 Pro or MiniMax H3?
If quality is the reason to move up from LTX 2.5 Fast, MiniMax H3 at 768P was the better step in our tests, and it costs less per second than LTX 2.5 Pro at 720p. We ran LTX 2.5 Pro on the same dialogue prompt: it framed the faces closer and the cook is more expressive, but it also sat the two of them at one table, as Fast did, while H3 kept the counter. LTX 2.5 Pro costs $0.108 per second at 720p and took 65 seconds; H3 costs $0.097 at 768P and took 4.1 minutes. We did not run Pro on the action or product prompts, so this rests on one scene.
What LTX's own benchmark says about MiniMax H3
LTX publishes its own comparison that includes MiniMax H3, and on quality our results point the other way. In its speed test, a 10-second image-to-video clip took 23.7 seconds through the LTX API at 1080p, against 180 seconds for MiniMax H3 at 768p measured on a third-party hosting service, queue included. In a separate check of 98 text-to-video prompts graded automatically for visible artifacts, where lower is better, LTX reports 0.28 for LTX 2.5 Pro, 0.39 for LTX 2.5 Fast and 0.46 for MiniMax, and calls the results preliminary. These are vendor-published figures, not our tests.
On speed our gap was of a similar size: LTX 2.5 Fast finished 4.5 to 9 times sooner than MiniMax H3. On quality, H3's clips looked more polished and followed the prompts more closely in all three of our scenes. Both findings can hold at once, because an automatic artifact score counts visible defects and does not check whether the shot is the one you asked for. Three prompts are a small sample, so run your own before you choose.
What each clip costs and how long it takes
On apimodels.app, LTX 2.5 Fast costs $0.081 per second at 720p, $0.117 at 1080p and $0.171 at 1440p, which is 10% below LTX's own API prices; LTX 2.5 Pro costs $0.108 at 720p and $0.153 at 1080p. MiniMax H3 costs $0.097 per second at 768P and $0.145 at native 2K. H3 renders at 2K unless you ask for 768P, so set the resolution on purpose. Failed generations are not charged. Prices as of 20 November 2026.
A budget for 100 six-second clips: LTX 2.5 Fast at 720p comes to 100 × 6 × $0.081 = $48.60, MiniMax H3 at 768P to 100 × 6 × $0.097 = $58.20, and LTX 2.5 Pro at 720p to $64.80. At each model's top tier, a 6-second clip costs $0.87 on H3 at native 2K and $1.03 on LTX 2.5 Fast at 1440p. Waiting times in our tests were 27 to 73 seconds per clip for LTX 2.5 Fast, 65 to 104 seconds for LTX 2.5 Pro and 4.1 to 4.8 minutes for MiniMax H3, queue included.
When neither model is the right choice
- You need 4K: neither, through our API. LTX advertises native 4K for its self-hosted weights; on our platform LTX 2.5 stops at 1440p and H3 at 2K.
- You need one clip longer than 20 seconds: neither. Generate shorter clips and join them in an editor.
- You need lots of drafts for as little as possible: MiniMax H3 Lite runs the H3 model at $0.02 per second at 768P and takes a first and last frame, but faces and limbs break in fast live-action movement, so treat its output as a draft.
- You need a first and a last frame at full quality: Seedance 2.0 takes both, plus up to nine reference images; see our LTX 2.5 vs Seedance 2.0 test.
How to run this comparison yourself
You can repeat this LTX 2.5 vs MiniMax H3 test with one API key; only the model name and the resolution change between the two requests.
- Create an API key in the apimodels.app console. New accounts get $1 of credit, enough for two 6-second LTX 2.5 Fast clips at 720p or one MiniMax H3 clip at 768P.
- Send the same prompt to the video generation endpoint twice: once with the model ltx-2.5-fast at resolution 720p, once with minimax-h3 at 768p, keeping the length and aspect ratio the same.
- Poll the task until its state is completed, then download the file from resultUrls; result links are temporary, so save what you want to keep.
- Write prompts in the order things happen, put spoken lines in quotes and end with a sentence that starts with "Sound:" and lists what you want to hear. For H3, describe each shot in turn; for LTX 2.5, state positions plainly and repeat the ones that matter.
{
"model": "minimax-h3",
"prompt": "Static medium shot inside a tiny ramen shop at night ... Sound: broth simmering, light rain on the window, their two voices.",
"resolution": "768p",
"duration": 6,
"aspect_ratio": "16:9"
}Sources
- LTX, "Fastest AI Video Generation Models (2026)", ltx.io/blog/ai-video-model-speed-comparison, read 20 November 2026: the speed test and the artifact scores quoted above.
- LTX, "LTX Video Generation API Pricing", ltx.io/model/api/pricing, read 20 November 2026: LTX 2.5 Fast at $0.09, $0.13 and $0.19 per second for 720p, 1080p and 1440p, and LTX 2.5 Pro at $0.12 and $0.17 for 720p and 1080p.
- MiniMax, MiniMax-H3 model card, huggingface.co/MiniMaxAI/MiniMax-H3, read 20 November 2026: the open base model at 768p, the 2K stage not released, the licence name and the application form.
- MiniMax H3 inputs, lengths, aspect ratios and prices: our MiniMax H3 model page and API docs, checked 20 November 2026.
- All clips, timings and charges: our own runs on apimodels.app on 20 November 2026.
Related models
- LTX 2.5 Fast: text, image and audio to video at 720p to 1440p, from $0.081 per second.
- LTX 2.5 Pro: the higher-quality LTX 2.5 tier at 720p and 1080p, from $0.108 per second.
- MiniMax H3: multimodal references and native sound at 768P or native 2K, 5 to 15 seconds, from $0.097 per second.
- MiniMax H3 Lite: the H3 model as a draft tier at 480P and 768P, from $0.01 per second.
- Seedance 2.0: full multimodal input with first and last frame, 480p to 1080p, from $0.092 per second.
Perguntas frequentes
Is LTX 2.5 faster than MiniMax H3?
Yes, by a wide margin. In our test LTX 2.5 Fast returned 6-second 720p clips in 27 to 64 seconds, while MiniMax H3 took 4.1 to 4.8 minutes at 768P on the same prompts, queue included. LTX's own benchmark reports 23.7 seconds for a 10-second 1080p clip through its API, against 180 seconds for H3 at 768p.
Which has better quality, LTX 2.5 or MiniMax H3?
In our three-prompt test MiniMax H3 was closer to the prompt every time: it kept the cook behind the counter, cut three shots in the written order and gave the product clip a dark, ad-like look with clear ice cubes. LTX 2.5 Fast drifted on staging and on the product. H3 still wrote made-up text on signs and labels. Three prompts are a small sample, so test your own before committing.
How much does a 6-second clip cost on LTX 2.5 and MiniMax H3?
On apimodels.app a 6-second clip costs $0.486 on LTX 2.5 Fast at 720p, $0.648 on LTX 2.5 Pro at 720p and $0.582 on MiniMax H3 at 768P. At the top tier it is $0.87 on H3 at native 2K and $1.03 on LTX 2.5 Fast at 1440p. H3 renders at 2K if you leave the resolution out. Failed generations are not charged.
Are LTX 2.5 and MiniMax H3 open source?
Both have published model weights, each under its own licence. LTX says LTX 2.5 is free for organisations under $10 million in annual revenue. MiniMax's public H3 weights cover the 768p base model without the 2K stage, and its model card sends users in the US, EU, UK and South Korea to an application form. Read each licence before running the weights yourself.
Can I use LTX 2.5 and MiniMax H3 with one API key?
Yes. On apimodels.app both run through the same video endpoint with the same API key; you switch by changing the model name from ltx-2.5-fast to minimax-h3 and the resolution from 720p to 768p. Requests are asynchronous: you create a task, then poll it or receive a callback.