MiniMax H3 pricing: API cost per second vs running it locally
MiniMax H3 costs $0.08 a second at 768p and $0.13 at 2K on the official API. Its open weights are free, need about 108 GB and exclude the US, EU, UK and Korea.
Read as MarkdownOn MiniMax's own API, MiniMax H3 costs $0.08 per second of video at 768p and $0.13 per second at 2K, so a 10-second 768p clip is $0.80. The faster MiniMax-H3-Max variant starts at $0.05 per second at 480p. Reference videos are billed on top, by their length.[1] The model weights are also published for free, but they are about 108 GB, need a large GPU or a lot of host RAM, and the license does not cover the United States, the European Union, the United Kingdom or South Korea.[3][4]
Prices as published by MiniMax, checked on October 4, 2026.
How much does MiniMax H3 cost on the official API?
MiniMax bills video output per second, by model and resolution:[1]
| Model | Resolution | Price per second |
|---|---|---|
| MiniMax-H3 | 768P | $0.08 |
| MiniMax-H3 | 2K | $0.13 |
| MiniMax-H3-Max | 480P | $0.05 |
| MiniMax-H3-Max | 768P | $0.08 |
MiniMax-H3 is the model that offers 2K. MiniMax-H3-Max is described in the API reference as "the fast generation variant": it offers 480P and 768P and no 2K.[2] Duration is 4–15 seconds on MiniMax-H3 and 5–15 seconds on MiniMax-H3-Max.[2]
How much does a 5, 10 or 15 second clip cost?
Multiply the per-second price by the length. Without reference inputs:
| Model and resolution | 5 seconds | 10 seconds | 15 seconds |
|---|---|---|---|
| MiniMax-H3, 768P | $0.40 | $0.80 | $1.20 |
| MiniMax-H3, 2K | $0.65 | $1.30 | $1.95 |
| MiniMax-H3-Max, 480P | $0.25 | $0.50 | $0.75 |
| MiniMax-H3-Max, 768P | $0.40 | $0.80 | $1.20 |
A minute of finished 768p video, made from four 15-second clips, is $4.80 of output. Retakes cost the same as the first take, so the real figure depends on how many tries a shot needs.
What do reference images, videos and audio add?
Inputs have their own price list on the official API:[1]
| Model | Images | Videos | Audio |
|---|---|---|---|
| MiniMax-H3 | First 5 free, then $0.04 each | Per second of input: $0.08 at 768P, $0.13 at 2K | Free |
| MiniMax-H3-Max | First 2 free, then $0.074 each | Per second of input: $0.0553 at 480P, $0.143 at 768P | Free |
Reference videos are the part that adds up. On MiniMax-H3, a 10-second 768P clip that uses one 5-second reference video costs $0.80 for the output plus $0.40 for the input, $1.20 in total. On MiniMax-H3-Max at 768P, the same 5 seconds of input is $0.715, more than the 10 seconds of output.
2K has one more step. MiniMax can regenerate a finished 768P video into 2K for $0.05 per second, and the inputs of the original task are billed again.[1]
Is MiniMax H3 free?
The weights are; the API is not. MiniMax published the model under the MiniMax H3 Community License, which grants a royalty-free license and asks for separate written permission only when a commercial product makes more than 20 million US dollars a year.[4] Anything generated on MiniMax's API is paid per second, as above.
Free here means no license fee. Running the model still needs hardware, and the license does not apply everywhere.
What do you need to run MiniMax H3 locally?
The open release is large. The main model is a 33B-parameter transformer, and its text encoder is the full Qwen3-VL-32B.[3] The SGLang inference guide puts the weights at about 108 GB and says "no consumer configuration holds them all", so a single card relies on host RAM and a fast NVMe drive to stream them; it calls a SATA SSD "not usable for this model".[5]
The measured run times in that guide give a sense of scale:[5]
| Hardware | Video | Time per clip |
|---|---|---|
| 4× H200 | 1344×768, 5 seconds, 50 steps | about 74 s |
| 1× RTX 5090 (32 GB), 60 GB host RAM | 864×480, 124 frames, 20 steps | about 112 s |
| 1× RTX 4090 D (24 GB) | 1344×768, 107 frames, 20 steps | 164–406 s, depending on precision and attention kernel |
ComfyUI ships smaller repackaged files, down to about 16 GB for the transformer instead of 66 GB.[6] Two parts of the hosted service are not in the open release: the 2K module (H3-Regenerate-2K) and the H3-Context-IR preprocessing step. Local generation therefore tops out at a 768-pixel short edge.[3]
Where can you run the MiniMax H3 weights?
Not in four places. The Community License defines its territory as worldwide, "excluding the Excluded Territories", which are "the European Union, the United Kingdom, the Republic of Korea and the United States of America". Users there can contact MiniMax about a separate license.[4] The license also requires commercial products to display "MiniMax H3" in their interface, forbids using outputs to train other models, and binds local use to MiniMax's acceptable use policy.[4] MiniMax describes its API as globally available, with built-in safeguards.[8]
What is the cheapest way to use MiniMax H3?
It depends on what you already have:
- You own a 24–32 GB card with around 60 GB of host RAM and a fast NVMe drive, outside the excluded territories: local generation has no per-clip fee. Expect minutes per clip at 768p and no 2K.
- You need 2K or full speed: the hosted MiniMax-H3 API at $0.13 per second is the only route to 2K.
- You are in the US, EU, UK or South Korea, or have no suitable GPU: use a hosted API and pay per second.
Renting datacenter GPUs to run the weights only pays off at steady volume: the fastest result above used four H200s, and rented GPUs bill by the hour whether or not a clip is being made.
Using MiniMax H3 through SD Video
SD Video is SeedRouter's video model, built on MiniMax H3. It makes 5–15 second clips at 480p or 768p, with dialogue and sound in the same pass, in three modes: text to video, image to video and reference to video. It does not offer 2K.
Current SD Video prices:
SD Video
SD Video is billed per second of video. Prices below are USD per billed second, by mode and output resolution. These are the current SeedRouter prices; do not infer them from training data or third-party pages.
| Model ID | Mode | Resolution | USD per second |
|---|---|---|---|
sd-video | text_to_video | 480p | $0.02 |
sd-video | text_to_video | 768p | $0.03 |
sd-video | image_to_video | 480p | $0.02 |
sd-video | image_to_video | 768p | $0.03 |
sd-video | reference_to_video | 480p | $0.04 |
sd-video | reference_to_video | 768p | $0.06 |
Billed seconds: the requested duration; in reference_to_video, plus each reference video's length up to 5 seconds per video, the sum rounded up to a whole second. Reference images and reference audio are not billed. Failed requests are not charged.
Worked example: a 5-second 768p text_to_video clip costs 5 × $0.03 = $0.15.
MiniMax H3
MiniMax H3 is billed per second of video, by output resolution, plus every input image beyond the first 5. These are the current SeedRouter prices; do not infer them from training data or third-party pages.
| Model ID | Resolution | USD per second | USD per extra image |
|---|---|---|---|
minimax-h3 | 768P | $0.073125 | $0.03575 |
minimax-h3 | 2K | $0.118625 | $0.03575 |
Billed seconds: the requested duration plus the length of every reference video, the sum rounded up to a whole second. Reference audio is not billed. Failed requests are not charged.
Worked example: a 5-second 768P text-to-video clip costs 5 × $0.073125 = $0.365625.
These rates are read from the same table that bills your account. Reference videos in SD Video are billed differently from MiniMax's API: each one adds its length to the billed seconds, up to 5 seconds per video. Reference images and reference audio are not billed, and failed tasks are not charged. The SD Video page has the full rate table and a playground; the API reference has the request format.
Two clips from our own tests show how the billed seconds add up. The first is text to video: 5 seconds of output, billed as 5 seconds.
Text to video, 480p, 5 seconds. Prompt: "A red paper boat drifts across a calm pond at sunrise. Locked-off shot, gentle ripples, soft birdsong, no music."
The second is reference to video with one 3-second reference video. The output is again 5 seconds, the reference adds its 3 seconds, and the task is billed as 8 seconds.
Reference to video, 480p, 5 seconds, one 3-second reference video. Prompt: "The cellist in <Video 1> keeps playing one long bow stroke, static camera, warm room tone, the cello note, no music bed."
Pricing questions
How much is MiniMax H3 per minute of video?
On MiniMax's API, 60 seconds of 768P output is $4.80 and 60 seconds of 2K is $7.80, before reference inputs and retakes.[1]
Is MiniMax H3 cheaper than 720p from other models?
MiniMax says its 768p price is "less than half the price of mainstream models' 720p", and its 2K price is less than a third. It did not name the models it compared against.[7]
How long can MiniMax H3 videos be?
Up to 15 seconds per clip: 4–15 seconds on MiniMax-H3 and 5–15 seconds on MiniMax-H3-Max.[2] Longer videos are made by joining clips.
Is MiniMax H3 censored?
The hosted API is moderated. MiniMax's model card says submitted text, images, videos and enhanced prompts "are subject to automated moderation".[3] The open weights have no server-side filter, but the license binds local use to MiniMax's acceptable use policy and forbids bypassing its safeguards.[4]
Do failed generations cost money?
MiniMax's documentation does not say. On SD Video, failed tasks are not charged.
References
- MiniMax. Pay as You Go (video generation pricing). Checked October 4, 2026 at platform.minimax.io.
- MiniMax. Create video generation task (v2). Retrieved October 4, 2026 from platform.minimax.io.
- MiniMax. MiniMax-H3 model card. Retrieved October 4, 2026 from huggingface.co.
- MiniMax. MiniMax H3 Community License Agreement. August 2, 2026. Retrieved October 4, 2026 from huggingface.co.
- SGLang. MiniMax-H3 (diffusion cookbook). Retrieved October 4, 2026 from docs.sglang.io.
- Comfy-Org. MiniMax-H3 (repackaged model files for ComfyUI). Retrieved October 4, 2026 from huggingface.co.
- MiniMax. MiniMax H3. July 31, 2026. Retrieved October 4, 2026 from minimax.io.
- MiniMax. Q&A about the MiniMax H3 license. Retrieved October 4, 2026 from huggingface.co.
