MiniMax H3 is hiapi's native 2K multimodal video model, and the part that matters most for short-form content — TikTok, Reels, Shorts — is that it renders directly in vertical 9:16 without any letterboxing tricks. We ran a real text-to-video generation through the MiniMax H3 model page on hiapi and embedded the actual output below, alongside the exact prompt we used and a practical breakdown of its three generation modes.
Summary
- MiniMax H3 generates natively at 2K and renders clean vertical 9:16 output — no cropping or upscaling needed for TikTok/Reels/Shorts delivery.
- It supports three generation modes: text-to-video, image-to-video (with first/last-frame control), and reference-driven generation from images, video, or audio.
- Clips run 4 to 15 seconds at $0.1189/second — a 4-second clip costs roughly $0.48, a 15-second clip roughly $1.78.
- Specific, physical prompt language (materials, light direction, camera framing) produced a clean, on-model result in our test — vague prompts are the most common cause of drifting or generic-looking output.
- All requests go through hiapi's single asynchronous task endpoint, the same submit-and-poll shape used across every video model on the platform.
What MiniMax H3 Actually Is
MiniMax H3 is a native 2K multimodal video model built for text-to-video, first/last-frame image-to-video control, and reference-driven generation using images, video, and audio, across clip lengths of 4 to 15 seconds. For short-form content specifically, the native 2K output matters more than it might seem: most vertical social formats top out around 1080×1920, so a 2K source gives you headroom to crop for a thumbnail or a slightly tighter reframe without visible quality loss — you're downscaling, not upscaling.
Text-to-Video: A Real Clip From a Real Prompt
We generated this clip directly from a text prompt, no reference image or video involved. Generate with this prompt → Select 2K and 9:16 in the Playground. Here's the actual output:
Prompt:
A barista's hand pours steamed milk from a stainless steel pitcher into a black espresso cup held at a slight angle on a light grey marble countertop. The milk stream settles and blooms into a full rosetta (fern-leaf) latte-art pattern by the end of the pour. Warm, directional side light rakes across the cup and pitcher. Vertical 9:16 framing, macro-close perspective, shallow depth of field, no camera movement, no text, no logos, no additional hands or objects entering frame.
Generate with this prompt → Select 2K and 9:16.
This is the kind of shot that works well for short-form: a single, physically grounded action (a pour, a reveal, a transformation) shot close and vertical, with enough specific detail — the material of the pitcher, the direction of the light, the exact pattern forming — that the model has almost nothing left to guess at. Naming the camera behavior explicitly (we asked for "no camera movement") is also worth doing even when you want a static shot; without it, models will sometimes add a slow, unrequested drift. For more shot ideas, see our MiniMax H3 prompt guide.
Image-to-Video: Starting From a Locked Frame
MiniMax H3's image-to-video mode takes an opening image through first_frame_image; add last_frame_image to guide the closing composition. A last frame requires a first frame, and these fields cannot be combined with image_urls, video_urls, or audio_urls. The input frame determines the video's aspect ratio. This is the better choice for short-form work when the opening frame is the hook — a product shot, a face, a specific piece of packaging — and you need the video to start from that exact composition rather than let the model interpret it from a text description. Text-to-video is faster to iterate on wording; image-to-video is what you reach for once you've already nailed the still frame you want the clip to open on.
Reference-Driven Generation: Images, Video, and Audio Together
The third mode is reference-driven generation, where MiniMax H3 can take cues from reference images through image_urls, videos through video_urls, and audio through audio_urls — rather than generating from a text prompt alone. In practice this is the mode to reach for when you already have an asset whose visual style, motion, or timing you want to carry into a new clip: matching a new subject to an established brand look, or syncing a generated clip's pacing to an existing audio track for a short-form cut. Audio references require at least one reference image or video. The limits are 5 images, 3 videos (each 2–15 seconds, 15 seconds total), and 3 audio clips (15 seconds total). It's a heavier, more deliberate workflow than text-to-video, so it's worth prototyping the shot with a plain text prompt first and only moving to reference-driven generation once you know what you're trying to match.
How Requests Work
Every video model on hiapi, including MiniMax H3, goes through the same asynchronous task flow: you submit a generation request and get back a task ID, then poll that task until it reaches success or fail. On success, select the type: "video" item in data.output and download its URL; on fail, inspect the returned error. There's no long-lived connection to manage and no model-specific request shape to relearn — the same submit-and-poll pattern in HiAPI's async task documentation covers every model on the platform, so switching from testing one video model to another is a matter of changing the model name and its input fields, not rebuilding your integration.
Pricing and Budgeting for Short-Form Batches
MiniMax H3 is billed at $0.1189 per second of output (see HiAPI's pricing page), with clip lengths from 4 to 15 seconds. That puts a single clip at roughly:
- 4 seconds: ~$0.48
- 8 seconds: ~$0.95
- 15 seconds: ~$1.78
For short-form content, a "shoot short, cut together" approach gives you more cut points and lets you regenerate one shot without replacing the whole sequence. Cost depends on total generated seconds: shorter shots save money only when their combined duration is lower. Count every successful candidate clip, including shots you later discard, in the budget.
FAQ
Does MiniMax H3 output true vertical video, or does it need cropping for TikTok/Reels? It renders natively at 2K and accepts a 9:16 aspect ratio directly — the clip above was generated vertical from the start, no cropping or reframing needed for short-form platforms.
How long can a MiniMax H3 clip be? 4 to 15 seconds per generation. For short-form content, stitching several shorter clips (4-6 seconds each) usually produces a more watchable result than one long single-shot render.
Can I start a MiniMax H3 video from my own image?
Yes — pass the opening image through first_frame_image, and optionally add last_frame_image. A last frame requires a first frame. Keep these separate from the reference fields image_urls, video_urls, and audio_urls.
What's the difference between image-to-video and reference-driven generation on MiniMax H3?
Image-to-video locks a starting (and optionally ending) frame from a still image. Reference-driven generation uses image_urls, video_urls, and audio_urls to guide style, motion, or timing; audio needs an accompanying image or video — useful when you're matching a new clip to an existing asset rather than starting from a blank frame.
How much does a typical short-form clip cost? At $0.1189/second, a 4-second clip runs about $0.48 and an 8-second clip about $0.95 (see HiAPI's live pricing page). Compare the total seconds of all successful clips, including regenerated shots; splitting a sequence alone does not reduce its price.
Is MiniMax H3 a good fit if I'm already using another hiapi video model, like Hailuo, for short-form clips? See our MiniMax H3, Seedance 2.0, and Hailuo 2.3 comparison, and pick based on the mode your shot needs: plain text-to-video, frame control, or multimodal references.
Takeaways
- MiniMax H3 renders natively in 2K vertical, so short-form clips need no cropping or upscaling for TikTok/Reels/Shorts.
- Specific, physically-grounded prompts (materials, light direction, explicit camera behavior) produced a clean result in our real test — vague prompts leave too much for the model to guess.
- Use text-to-video to iterate on an idea quickly, image-to-video when the opening frame is locked in advance, and reference-driven generation when you're matching an existing image, video, or audio asset.
- At $0.1189/second across a 4-15s range, budget by the total seconds of successful candidate clips; shorter shots make selective regeneration easier.
- All requests go through the same task-submit-and-poll flow as every other video model on hiapi, so testing MiniMax H3 alongside models like Hailuo 2.3 doesn't require separate integration work.
Generate with this prompt → Start with the latte-art shot in 2K, 9:16.
