How many seconds MiniMax H3 really generates by resolution

MiniMax H3 advertises "up to 15 seconds". That number is true and also misleading, because the model does not render on whole seconds. It renders on a frame grid, and most requested durations snap to a fractional result. A 10-second request comes back as 10.125 seconds. A 14-second request lands at 13.667.

If you cut to music, stitch clips end to end, or budget API costs, the grid matters more than the marketing number. This guide maps every requested duration from 4 to 15 seconds to what H3 actually renders, at both supported resolutions, 768P and 2K.

TL;DR

• H3 generates 4-15 second clips, whole numbers only. 3 seconds is rejected, 7.5 is rejected MiniMax API video generation docs.
• Output is 24 fps, and the frame count follows frames = 17n + 5. Only one value in the window lands on a whole second: 8 seconds (192 frames) minimax-h3-spec validation rules.
• There is no 1080p and no 4K. The API accepts 768P and 2K and rejects anything else, it does not downgrade minimax-h3-spec validation rules.
• 2K is not an upscale. The 768P result is fed back through the model with the original context and regenerated, which is why small on-screen text survives minimax-h3-spec validation rules.
• Billing follows rendered seconds, not requested ones. A 10-second 2K request bills as 10.125 seconds, about $1.32 at list price minimax-h3-spec validation rules.

The duration window

The API reference sets the window at 4 to 15 seconds, integer values only. duration is required, has no default, and the check is strict: 7.5 fails, the string "8" fails minimax-h3-spec validation rules. Many write-ups say the floor is 5 seconds. The API accepts 4, and the spec records that discrepancy minimax-h3-spec validation rules.

The H3-Max variant, the post-trained model MiniMax released with fal.ai, runs 5 to 15 seconds MiniMax API video generation docs. Both variants output at 24 fps with native 32 kHz stereo audio MiniMax H3 GitHub README.

The frame grid: why 10 seconds is not 10 seconds

H3 renders in frame blocks. The frame count follows frames = 17n + 5, and the output runs at 24 fps. Your requested duration snaps to the nearest grid point minimax-h3-spec validation rules.

The table below is derived from that formula at 24 fps. Every requested second in the window, and the clip it actually produces:

Two consequences, both practical:

• Cutting to music or stitching clips? Use 8 seconds. It is the only whole second in range, and every other value drifts, with the drift accumulating across a sequence minimax-h3-spec validation rules.
• Budgeting? You are billed on rendered seconds. A 10-second request bills as 10.125 minimax-h3-spec validation rules.

The spec notes the 10-second case explicitly: duration 10s renders as 243 frames = 10.125s (n=14). The only whole second in range is 8.000s (192 frames) minimax-h3-spec validation rules.

What resolution actually changes

Resolution does not change the duration window. The same 4-15 second range applies at 768P and at 2K MiniMax API video generation docs. What resolution changes is the render path and the cost.

768P

The base checkpoint generates 768p. The short side is 768 pixels by default, and the supported dimensions follow the aspect ratio MiniMax H3 GitHub README. Runware documents the 768p dimensions: 1344x768 for 16:9, 1024x768 for 4:3, 768x768 for 1:1, 768x1024 for 3:4, and 768x1344 for 9:16 Runware MiniMax H3 API docs.

This is the tier that local H3 output tops out at. The open H3-Base checkpoint produces 768p, and 2K requires a separate hosted regeneration step MiniMax H3 model card.

2K

2K is the highest direct output tier, and it is not an upscale. The 768P result is fed back through the model together with your original context and generated again. That is why small on-screen text survives the jump, and why 2K costs more per second rather than a flat surcharge minimax-h3-spec validation rules.

The launch post frames 2K as the headline spec: "up to 15 seconds at 2K resolution" MiniMax H3 launch post. There is nothing between 768P and 2K, and nothing above 2K. A request for 1080p is rejected, not downgraded to the nearest tier minimax-h3-spec validation rules.

Aspect ratio

Six ratios are accepted: 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. No arbitrary dimensions minimax-h3-spec validation rules.

One trap: in image-to-video modes the output ratio follows the input image, not the aspect_ratio field. If you need a 9:16 clip from a 16:9 still, crop the still. Setting the field will not do it minimax-h3-spec validation rules.

Where quality degrades: length, not resolution

Resolution is a render-path decision, so it does not silently degrade. Duration does. Longer clips mean more frames for the model to keep consistent, and the prompt structure expects every detail to correspond to something visible in a specific shot MiniMax H3 GitHub README.

The practical pattern from the spec and the API docs:

• 4-8 seconds is the safe range for prompt adherence. The 8-second grid point is the cleanest cut target minimax-h3-spec validation rules.
• 9-15 seconds is where multi-shot prompts earn their keep. The prompt ceiling is 7,000 characters, which is generous, but a long clip with one flat description is the failure mode the structure is designed against minimax-h3-spec readthedocs rules.
• For anything longer than 15 seconds, you stitch. The 8-second grid point exists for exactly that minimax-h3-spec validation rules.

Cost per rendered second

The 2K 10-second figure matches the spec's own worked example: $1.3163 for 10.125 rendered seconds minimax-h3-spec validation rules. Reference images bill separately: first 5 free, then $0.04 each MiniMax API pricing overview.

For the full local vs hosted cost picture, including VRAM tiers and when cloud wins, see the MiniMax H3 VRAM and GPU requirements guide. The six-second contest rule maps to the shortest practical grid clips, covered in the H3 content contest guide. Local setup paths are in the MiniMax H3 ComfyUI guide.

Common errors and fixes

FAQ

How long can MiniMax H3 generate? Up to 15 seconds in a single generation, whole seconds only. The frame grid means most requested values render slightly longer or shorter, and only 8 seconds lands exactly minimax-h3-spec validation rules, MiniMax API video generation docs.

How many seconds can MiniMax H3 do? 4 to 15 per clip, 24 fps. The window is inclusive on both ends, and duration is a required integer field MiniMax API video generation docs.

What is the best resolution for MiniMax H3 video generation? 768P for speed and cost, 2K for final output where text and detail matter. 2K is a regeneration of the 768P result with the original context, not an upscale, so it costs more per second minimax-h3-spec validation rules.

Can MiniMax H3 do i2v? Yes. One input image becomes a first-frame or last-frame video, and two images become a first-and-last-frame clip. The output aspect ratio follows the input image MiniMax H3 GitHub README, minimax-h3-spec validation rules.

Why is my 10-second clip 10.125 seconds? The frame count follows frames = 17n + 5 at 24 fps, and the request snaps to the nearest grid point. 243 frames is the closest to 10 seconds minimax-h3-spec validation rules.

Does 2K change the maximum duration? No. The 4-15 second window is identical at both resolutions; resolution changes the render path and the per-second price MiniMax API video generation docs.

Sources

• MiniMax API: Video Generation — Duration window 4-15s (H3) and 5-15s (H3-Max), resolution tiers, reference input limits.
• minimax-h3-spec: Validation rules — Frame grid formula (17n+5 at 24 fps), 10s renders as 10.125s, 8s is the only whole second, 1080p/4K rejected, 2K regeneration, aspect ratio rules, billing on rendered seconds.
• MiniMax-AI/MiniMax-H3 (GitHub) — Output specs (24 fps, 32 kHz stereo, 11 dialogue languages), FL2VA/Ref2VA variants, prompt writing skill.
• MiniMaxAI/MiniMax-H3 (Hugging Face) — Model card: open checkpoints, 768p base output, 2K via hosted regeneration.
• MiniMax H3 launch post — Official specs: up to 15 seconds at 2K, native stereo audio.
• Runware MiniMax H3 API docs — 768p pixel dimensions per aspect ratio, per-second pricing.