Table of Contents

  1. MiniMax Open-Weight Strategy Overview
  2. MiniMax Open-Weight Model Lineup (M1 · M3 · H3)
  3. H3 Core Technologies
  4. Getting Started — API
  5. Getting Started — Local Deployment (Open Weight)
  6. Usage Examples — T2V · I2V · Ref2V
  7. Strengths
  8. Weaknesses
  9. Impact on the Video Production & Editing Industry
  10. Key Competitors
  11. Summary

1. MiniMax Open-Weight Strategy Overview

Following its text-centric LLMs (M series), MiniMax announced H3, its first open-weight video generation model, in July 2026, extending its open-weight strategy to the full image, audio, and video stack.

Key strategic directions:

  • Modality unification: A single omni-modal model that understands text, image, video, and audio as one input context and jointly generates video with stereo audio.
  • Task unification: Tasks previously split across specialized models — text-to-video (T2V), image-to-video (I2V), reference-to-video (Ref2V), video editing, and voice/SFX/music generation — are generalized into a single model.
  • Open-weight openness: The latest H3 weights are publicly released, so you can deploy locally or in a private environment, use it without API fees, or fine-tune it for specific domains.
  • Price disruption: Per-second pricing at 2K is less than a third of mainstream competitors; at 768p it is less than half the price of mainstream 720p output.

2. MiniMax Open-Weight Model Lineup

Model Released Key features & performance
MiniMax-M1 June 2025 World's first open-weight hybrid-attention reasoning model
- Architecture: hybrid MoE with 456B total parameters and 45.9B active parameters per token
- Context: up to 1M tokens (8x DeepSeek-R1)
- Efficiency: only 25% of DeepSeek-R1's FLOPs to generate 100K tokens
- Benchmarks: outperforms leading open models including DeepSeek-R1 and Qwen3-235B
MiniMax M3 June 2026 First open-weight model combining coding, 1M context, and native multimodality
- Architecture: MoE with ~428B total parameters and ~23B active parameters per token
- Coding: 59.0% on SWE-Bench Pro (beats GPT-5.5 at 58.6%)
- Price: API priced at 5-10% of competing closed models ($0.3/M input tokens during promotions)
MiniMax H3 July 2026 First MiniMax video model released as open weight
- Performance: #1 in video editing, #2 in text-to-video
- Capability: takes text, images, video, and audio as input, generating up to 2K-resolution, 15-second videos with native stereo audio
- License: MiniMax Community License (commercial use permitted for organizations with annual revenue under $20M)
- Open weight: HuggingFace MiniMaxAI/MiniMax-H3 (33B, BF16)

3. H3 Core Technologies

H3 consists of three modules. The API provides all three in one integrated call; the open-weight release includes only H3-Base.

Module Role Availability
H3-Context-IR Understands and refines multimodal input, converting it into a structured prompt (Intermediate Representation). Uses language as the "bridge for generalization" API only
H3-Base Generates video + stereo audio at 768p from the Context-IR output (33B dense Omni-Transformer) Open weight
H3-Regenerate-2K Re-injects the original context with the 768p output to regenerate at 2K. Reuses the base model's generative capability instead of a dedicated super-resolution module API only

Key technical highlights:

  • H3-Contextual Omni Representation: A captioning pipeline covering video, audio, and multi-shot relationships. Most source material is distilled from roughly 100K tokens of inference down to an average of ~4K tokens.
  • H3-VAE: A complete tokenizer overhaul that improves reconstruction quality and learnability at once. Its high compression ratio delivers a 4x gain in effective sequence length — the key technology behind native 2K support.
  • H3-Omni Transformer: A dense Transformer built for task generalization. It separates understanding and generation workloads to optimize hardware utilization, lifting training throughput by nearly 30%. MM-RoPE captures (time, height, width) 3D positional relationships, and native sparse attention is supported.
  • In-Context Regeneration: Instead of a dedicated super-resolution module, H3's base model regenerates its own low-resolution output together with the original multimodal context, recovering small text and fine details that conventional SR can only "guess."

H3 Input/Output Specification

Item Spec
Output duration 4-15 seconds
Output resolution 768p (default) up to 2K (via H3-Regenerate-2K)
Output frame rate 24 FPS
Output audio 32 kHz stereo
Aspect ratios 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, adaptive
Dialogue languages Stable support for 11 languages incl. Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish

4. Getting Started — API

H3 uses an asynchronous generation flow. You submit a generation request, receive a task_id immediately, then poll the task status and download the finished video.

4.1 Prerequisites

  1. Sign up at platform.minimax.io
  2. Create an API key under Account Management > API Keys (used as the Authorization: Bearer <API_KEY> header)
  3. Top up balance/token plan (2K per-second pricing is under a third of mainstream models)
export MINIMAX_API_KEY="<YOUR_API_KEY>"
export MINIMAX_API_BASE="https://api.minimax.io"   # Global (CN: https://api.minimaxi.com)

4.2 Create Task

POST {MINIMAX_API_BASE}/v2/video_generation

curl -X POST "$MINIMAX_API_BASE/v2/video_generation" \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "content": [
      {
        "type": "text",
        "text": "Epic space-opera theatrical teaser: a female captain stands alone before a massive observation window as the last fleet gathers and jumps away in a blinding flash, the bridge shaking, leaving her behind."
      }
    ],
    "resolution": "2K",
    "duration": 5,
    "ratio": "16:9"
  }'

You receive a task_id in the response.

{ "task_id": "424010985738629" }

4.3 Query Task

GET {MINIMAX_API_BASE}/v2/query/video_generation/{task_id}

curl -s "$MINIMAX_API_BASE/v2/query/video_generation/424010985738629" \
  -H "Authorization: Bearer $MINIMAX_API_KEY"

On success, status becomes succeeded and content.url returns the video download URL.

{
  "task": {
    "id": "424010985738629",
    "model": "MiniMax-H3",
    "status": "succeeded",
    "content": { "url": "https://your-cdn.example.com/h3-generated-2k-output.mp4" },
    "resolution": "2K",
    "duration": 5,
    "ratio": "16:9"
  }
}
  • Statuses: queued → running → succeeded / failed / cancelled
  • Tasks can only be queried within the last 7 days, and video URLs are time-limited — download and store them promptly.
  • On failure, an error field with code/message is returned.

4.4 Complete Python Example (create → poll → download)

import time
import requests

API_BASE = "https://api.minimax.io"
API_KEY = "<YOUR_API_KEY>"
HEADERS = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

def create_video(prompt, resolution="2K", duration=5, ratio="16:9"):
    resp = requests.post(
        f"{API_BASE}/v2/video_generation",
        headers=HEADERS,
        json={
            "model": "MiniMax-H3",
            "content": [{"type": "text", "text": prompt}],
            "resolution": resolution,
            "duration": duration,
            "ratio": ratio,
        },
    )
    resp.raise_for_status()
    return resp.json()["task_id"]

def wait_video(task_id, timeout=600):
    url = f"{API_BASE}/v2/query/video_generation/{task_id}"
    deadline = time.time() + timeout
    while time.time() < deadline:
        data = requests.get(url, headers=HEADERS).json()["task"]
        if data["status"] in ("succeeded", "failed", "cancelled"):
            return data
        time.sleep(5)
    raise TimeoutError("task did not finish in time")

task_id = create_video("A boy playing basketball by the sea, golden hour, cinematic")
result = wait_video(task_id)
if result["status"] == "succeeded":
    video_url = result["content"]["url"]
    mp4 = requests.get(video_url).content
    open("output.mp4", "wb").write(mp4)
    print("saved output.mp4")

4.5 Request Parameter Summary

Parameter Type Required Description
model string yes MiniMax-H3 (currently the only one)
content array yes Multimodal input array with text/image_url/video_url/audio_url elements. Every request must include one non-empty text (prompt) (max 7000 characters)
resolution string yes 768P or 2K
duration int yes 4-15 (seconds)
ratio string conditional adaptive (default), 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. T2V (text only) cannot use adaptive and must specify one; image input is always treated as adaptive
callback_url string no POST notification on status change. Must answer the challenge verification within 3 seconds

Input media limits (total request body ≤ 64MB; use public URLs for large files):

Media Limits
Image (image_url) JPG, JPEG, PNG, WEBP, HEIC, HEIF; ≤ 30MB each; [256, 5760]px; aspect ratio [0.4, 2.5]. first frame ≤ 1, last frame ≤ 1, reference images ≤ 9
Video (video_url, reference scenario) MP4, MOV; H.264/H.265, AAC, MP3; ≤ 50MB each; ≤ 3 clips; 2-15s per clip, total ≤ 15s
Audio (audio_url, reference scenario) WAV, MP3; ≤ 15MB each; ≤ 3 clips; 2-15s per clip, total ≤ 15s (audio alone is not allowed)

5. Getting Started — Local Deployment (Open Weight)

5.1 Download the Model

Get MiniMaxAI/MiniMax-H3 (33B params, Image-Text-to-Video, BF16) from HuggingFace. It is released as two task-specific checkpoints: FL2VA (first/last-frame) and Ref2VA (reference-based).

# Original checkpoints (for SGLang, vLLM) — both task families
hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" "Ref2VA/*" --local-dir MiniMax-H3

# A single task family
hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" --local-dir MiniMax-H3

diffusers users need no manual download: ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3") fetches exactly the components it needs.

5.2 Recommended Inference Frameworks

Framework Notes
SGLang 4-GPU recommended, Ulysses parallelism for high performance. Official cookbook available
vLLM Optimized cache & scheduling; vllm recipes provided
diffusers One-line ModularPipeline.from_pretrained loading, Python pipeline
ComfyUI Comfy-Org/MiniMax-H3 for node-based workflows. T2V/R2V templates provided

5.3 SGLang Serving Example

# FL2VA (text/first-last frame → video)
sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010 \
  --model-variant fl2va

# Ref2VA (reference image/video/audio → video)
sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30011 \
  --model-variant ref2va

5.4 Full 2K Workflow (API + Local Hybrid)

The open weights reproduce 768p output locally; 2K quality can be reproduced by combining them with the official API.

  1. H3-Context-IR (POST /v2/h3_context_ir): convert input into a structured prompt before local generation. Critical to final quality — strongly recommended.
  2. H3-Base (local SGLang): generate 768p video from the converted prompt.
  3. H3-Regenerate-2K (POST /v2/video_regeneration): regenerate 768p output with the original context to 2K.
# Environment variables
export SGLANG_DEPLOYMENT_URL="<sglang-deployment-url>"
export MINIMAX_API_BASE="https://api.minimax.io"   # or https://api.minimaxi.com (CN)
export TOKEN="<token>"

5.5 License

H3 is distributed under the MiniMax H3 Community License Agreement.

  • Organizations with annual revenue of under $20M may use it commercially.
  • Weights can be downloaded, modified, and fine-tuned.
  • Organizations above the threshold need a separate commercial license.
  • Official license file: LICENSE in huggingface.co/MiniMaxAI/MiniMax-H3.

6. Usage Examples — T2V · I2V · Ref2V

H3 supports several generation scenarios based on the content array combination.

6.1 Text-to-Video (T2V)

{
  "model": "MiniMax-H3",
  "content": [
    { "type": "text", "text": "A cinematic product commercial of a luxury watch floating above a city at dawn, dramatic lighting" }
  ],
  "resolution": "2K",
  "duration": 5,
  "ratio": "16:9"
}

6.2 Image-to-Video (I2V) — First/Last Frame

{
  "model": "MiniMax-H3",
  "content": [
    { "type": "text", "text": "Pull focus to the people in the background and add more steam to the ramen bowl." },
    { "type": "image_url", "image_url": { "url": "https://example.com/first_frame.png" }, "role": "first_frame" },
    { "type": "image_url", "image_url": { "url": "https://example.com/last_frame.png" }, "role": "last_frame" }
  ],
  "resolution": "2K",
  "duration": 5,
  "ratio": "adaptive"
}

6.3 Reference-to-Video (Ref2V) — Video Editing, Voice & Music

{
  "model": "MiniMax-H3",
  "content": [
    {
      "type": "text",
      "text": "Character speaks: Follow the wind, live free. Voice timbre follows reference audio 1."
    },
    { "type": "video_url", "video_url": { "url": "https://example.com/source.mp4" }, "role": "reference_video" },
    { "type": "audio_url", "audio_url": { "url": "https://example.com/voice.mp3" }, "role": "reference_audio" }
  ],
  "resolution": "2K",
  "duration": 5,
  "ratio": "adaptive"
}

6.4 Role Combination Rules

Scenario content composition
T2V text only
I2V first frame text + 1 image (first_frame, or role omitted)
I2V last frame text + 1 image (last_frame)
I2V first & last frame text + 2 images (first_frame + last_frame)
Ref2V text + any combination of reference image (reference_image), video (reference_video), audio (reference_audio); audio alone not allowed, at least one image/video required

Note: I2V and Ref2V are mutually exclusive. If any reference_* role appears, first_frame/last_frame must not be used.

7. Strengths

  • Open-weight freedom: Download the weights and deploy in a local or private environment. Strengthen data security, use it without per-call API fees, or fine-tune for specific domains.
  • Frontier-level performance: #1 in video editing, #2 in text-to-video. Mature enough to be put straight into commercial content production (ads, branding, e-commerce, gaming).
  • Omni-modal input: Takes text, images, video, and audio together and executes compound instructions in natural language, e.g. "reference the camera movement of this video, have the character in that image sing with this voice."
  • Native stereo audio: Generates 32kHz stereo sound along with video — no separate TTS, music, or SFX pipeline. Audio and video are synchronized in one pass.
  • Disruptive cost efficiency: 2K per-second price under a third of mainstream models; 768p under half of mainstream 720p. Consistent with the M3 LLM price strategy (5-10% of closed models).
  • Long segments: Single generation of up to 15 seconds plus multi-clip references enable natural multi-shot direction.
  • Fast ecosystem adoption: Official ComfyUI integration and multiple HuggingFace Spaces demos (multimodal, Ref2VA, FL2VA, etc.) appeared immediately after launch.

8. Weaknesses

  • Hardware requirements: Running the 33B dense model smoothly locally requires expensive infrastructure such as high-end GPUs (recommended 4-GPU setup).
  • Some modules are not open: H3-Context-IR and H3-Regenerate-2K are API-only. To fully reproduce top quality (especially 2K), even local deployments must be combined with the official API.
  • License restrictions: The MiniMax Community License limits commercial use by organizations with annual revenue over $20M.
  • Limited independent verification: Announced benchmarks (e.g., #1 video editing) are MiniMax's own measurements and need independent verification.
  • It's a Chinese company: Data security and regulatory concerns may constrain usage in some countries and enterprises.
  • Irregular generation quality: Faces, hands, and fine text still have limits; prompt tuning and multiple regenerations (seeds) are often required.
  • Regulatory & copyright risks: Portrait rights, brand logos, music copyright, and deepfake abuse make H3 subject to generative-media regulations.

9. Impact on the Video Production & Editing Industry

H3 has the potential to change the entire video production pipeline, not just "generate a clip." The #1 video editing ranking is especially significant.

9.1 Collapse in Production Cost

  • 2K per-second cost is under a third of mainstream models. The direct generation cost of a 30-second brand video could fall to 1/10 or less of its former level.
  • With open weights + local deployment, per-second billing disappears entirely. With GPUs, video marginal cost drops to the level of electricity.
  • Combined with the M3-style "5-10% of closed-model price" strategy, the price floor of the whole generative media market could be reset.

9.2 Editing Paradigm Shift: From "Manipulation" to "Description"

Traditional editing means trimming, joining, and applying filters on a timeline. H3 regenerates video and audio wholesale from a natural-language description of the desired result.

  • V2V editing that only swaps characters, backgrounds, or mood of a scene
  • Compound editing that borrows the camera movement of a reference video and applies the voice timbre of a reference audio
  • Frame-anchored editing that directs "middle scenes" by specifying first and last frames
  • Shot planning that stitches 15-second multi-shots from 4-second audio/video references

In other words, the paradigm shifts from "cut-based editing" to "intent-based generation." Traditional NLE tools like Adobe Premiere and DaVinci Resolve face pressure as some of their core value gets absorbed into generation engines.

9.3 Democratization of Producers

  • With only a direction/scenario idea, a solo creator or a small team can produce ad, brand, e-commerce, and product demo videos.
  • Things that previously required outsourcing — moving product-website banners, animated posters, opening titles, UI/UX demos — can be produced instantly from a prompt.
  • The language barrier also drops. 11-language dialogue support lets you generate localized videos without separate casting.

9.4 Workflow Shift: From "Clip Generation" to "Full Production Participation"

As MiniMax's roadmap states, video models will move beyond single-clip generators to participate in the entire content production process.

  • Pre-viz: scenario → storyboard video → direction validation before shooting
  • Location/background/CG replacement: replace live shoots with generated backgrounds (product videos, interior, architectural presentations)
  • Automatic voice/music/SFX: synthesize dialogue, narration, BGM, and SFX in one pass from references
  • Versioning & A/B testing: generate dozens of marketing creatives and compare performance

9.5 Industry-by-Industry Impact Summary

Industry Impact
Advertising & marketing Creative A/B testing costs collapse; brand consistency becomes easier (brand-rendering accuracy is a strength)
E-commerce Product videos, 360 demos, lifestyle cuts generated from a single photo
Gaming Cinematic trailers, character-motion references, concept videos produced quickly
Film & TV Pre-viz, production design, VFX assistance. Beware legal risks around face/person reproduction
Broadcasting & YouTube Solo media creators produce high-quality backgrounds, motion thumbnails, and openings directly
Education & research Demo/explainer videos generated on the fly for learning and presentations

9.6 Risks & Responses

  • Employment shifts: Automation pressure will start with lower-end editing and motion-graphics jobs. Conversely, new roles emerge for reviewing, curating, and legally validating generated output.
  • Deepfakes & fraud: Reference-based face/voice editing becomes easy, raising the importance of watermarking and detection technologies.
  • Copyright & portrait rights: Rights over reference media and generated output are ambiguous. Contract and license management becomes essential.
  • Media literacy: Labeling regulations for "generated video" are expected to tighten.

In conclusion, H3-class models lower the cost floor of video production, democratize who can produce, and move the paradigm of editing from "editing" to "generation" — acting as a catalyst for structural change in the industry.

10. Key Competitors

MiniMax H3 competes with the following strong rivals.

  • ByteDance (Seedance): A direct competitor in the video-AI space.
  • Google (Veo · Gemini): A major force in text-to-video and multimodal generation. Gemini 3.1 Pro and Gemini Omni Flash are strong rivals to M3 and H3.
  • OpenAI (Sora): Top tier in text-to-video. Leads the generative media market since Sora 2.
  • Runway: Strengths in cinematic editing and control (Gen-4, etc.).
  • Kling (Kuaishou): A leading Chinese video generation player.
  • Other Chinese open-weight camps: In text LLMs, DeepSeek (R1, V4-pro), Qwen (Alibaba, Qwen3-235B, Qwen3.7-Max), and Moonshot AI (Kimi) compete with M1/M3; some are expanding into video models.

11. Summary

MiniMax has extended its open-weight strategy to omni-modal media with H3 (video generation), following M1/M3 (text reasoning and coding). Frontier performance (#1 video editing), generation of up to 2K, 15-second videos with stereo audio, pricing under a third of mainstream rivals, and a community license open to organizations under $20M in annual revenue — H3 is a signal that the open-weight ecosystem is entering a market long dominated by closed video models.

In the video production and editing industry specifically, H3 (1) lowers production cost from per-second billing down to GPU-level marginal cost, (2) shifts the paradigm toward natural-language, intent-based editing, and (3) accelerates the democratization of solo and small-team production. As long as you account for independent benchmark verification and license/regulatory risks, H3 is a realistic adoption candidate across a wide range of video workflows — advertising, e-commerce, gaming, and broadcasting.

References