MiniMax H3 API: Complete Pricing & Integration Guide

Key Specifications

API Pricing Breakdown

MiniMax H3 uses per-second billing for video output. The Context-IR preprocessing step is billed per token.

Video Generation

Reference Inputs

Price Per Clip

Note: 768P → 2K regeneration at $0.05/sec means drafting at 768P then upgrading costs exactly the same as shooting 2K directly ($0.08 + $0.05 = $0.13). The cheaper path only saves money on takes you discard.

Architecture Pipeline

MiniMax H3 uses a three-stage pipeline for 2K output:

The H3-Context-IR step is critical to output quality. MiniMax strongly recommends incorporating it into your generation pipeline rather than building a custom preprocessing system.

API Integration

Generate video via the Video Generation V2 endpoint:

curl -X POST https://api.minimax.io/v2/video_generation \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "prompt": "A cinematic aerial shot of a futuristic city at sunset",
    "resolution": "2K",
    "duration": 10,
    "aspect_ratio": "16:9"
  }'

Model ID must be MiniMax-H3 exactly (capitalization and hyphen matter). The endpoint returns a task ID; poll for completion or use the callback URL.

Open Weights

MiniMax released H3-Base checkpoints under the Community License on August 2, 2026:

Self-hosted output is limited to 768P. For 2K, you must use the hosted API's H3-Regenerate-2K pass or build your own super-resolution pipeline.

Download: Hugging Face — MiniMaxAI/MiniMax-H3 | GitHub

MiniMax H3 Max (fal.ai Variant)

H3 Max is fal.ai's post-trained variant of MiniMax H3, optimized for speed and prompt adherence. Key differences from base H3:

• 768P output (not 2K)
• Sub-3-second render time for 5-second clips
• Stronger prompt adherence and aesthetics
• 5 free generations per day for signed-in users
• Currently available exclusively on fal.ai
• Open weights planned by fal.ai

Use H3 for 2K output or reference/editing workflows; use H3 Max for rapid prototyping and high-volume generation at speed.

Comparison: MiniMax H3 vs Competitors

MiniMax H3 is the only model in this class that offers native stereo audio, 2K output, and open weights — at a lower per-second price than most competitors at lower resolutions.

Rate Limits & Constraints

Frequently Asked Questions

How much does 10 seconds of MiniMax H3 video cost?

$1.30 at 2K ($0.13/sec) or $0.80 at 768P ($0.08/sec). Reference images past the fifth add $0.04 each. Reference video input is billed at the output resolution rate.

Is MiniMax H3 cheaper than Kling or Runway?

Yes, significantly. MiniMax H3 at 2K ($0.13/sec) is cheaper than Kling at 1080P ($0.14/sec) and Runway Gen-4 ($0.15/sec), while offering higher resolution and native audio.

Can I self-host MiniMax H3?

The base model weights are open under the Community License, but self-hosting is limited to 768P output. The Context-IR preprocessing and 2K regeneration steps require the hosted API. The transformer alone is 61.7 GB in BF16.

Does MiniMax H3 support text generation?

No — MiniMax H3 is a video generation model. For text LLM APIs (chat, reasoning, code generation), consider DeepSeek, Qwen, Kimi, or GLM through BestLLM.