How to Use MiniMax H3 with AnyCap: A Complete Guide to AI Video Generation with Native Audio

MiniMax released H3 on July 31, 2026, and it's already available inside AnyCap as minimax-h3. It generates video with native stereo audio baked in, up to 15 seconds at 2K, from a single CLI call. No separate audio pass, no stitching clips together afterward. Below is how to actually run it, mode by mode, with real output from the model itself (not stock footage or mockups).

What Is MiniMax H3?

MiniMax H3 is a general-purpose, omni-modal generation model. Unlike earlier video models that split tasks into narrow specialists (text-to-image, image-to-video, motion reference, voice reference, and so on), H3 was trained to understand a unified multimodal context (text, images, video, and audio together) and generate directly from it.

Quick facts:

The model is built for commercial content work: advertising, branding, e-commerce, product design, UI/UX mockups, gaming assets, and film-style openers, rather than pure novelty clips.

Why Generate MiniMax H3 Videos Through AnyCap

You can call MiniMax H3 directly through MiniMax's own API. The case for routing it through AnyCap instead comes down to what happens before and after that one call.

Authenticate once with anycap login, and you're not just talking to H3: veo-3.1, sora-2-pro, kling-3.0, and seedance-2 sit under the same key, so testing a prompt across five models means changing one flag instead of managing five sets of credentials. Every video call also returns the same flat JSON shape (status, local_path, model, credits_used, request_id), so a script written against one model keeps working when you swap in another. And once a clip lands on disk, AnyCap's Drive and Page features turn it into a shareable link or a hosted page without a separate upload step.

CLI vs. MCP: Two Ways to Call MiniMax H3 Through AnyCap

AnyCap exposes MiniMax H3 through two different surfaces, and it's worth picking the right one before you start scripting anything.

Use the CLI when you're calling MiniMax H3 from a terminal, a Bash/Python pipeline, or anywhere you already control process invocation directly; it's the fastest path with no host configuration.

Use MCP when your agent (Claude Code, Cursor, Codex, Claude Desktop, etc.) speaks MCP natively and you want it to discover anycap_video_generate and anycap_model_schema as typed tools instead of shelling out to a subprocess. This also keeps file-system access explicitly scoped via --allow-root, which matters if the agent is operating semi-autonomously.

Setting up the MCP server

Requires AnyCap CLI v0.5.0 or later (anycap update if anycap mcp --help doesn't resolve).

Claude Code:

claude mcp add --scope local anycap -- anycap mcp --allow-root "$PWD"

Codex:

codex mcp add anycap -- anycap mcp

Generic stdio host (add to the host's MCP config file):

{
  "mcpServers": {
    "anycap": {
      "type": "stdio",
      "command": "anycap",
      "args": ["mcp", "--allow-root", "/absolute/path/to/media"]
    }
  }
}

Once connected, generating a MiniMax H3 clip through MCP is the same request shape as the CLI, just routed through the anycap_video_generate tool instead of a shell command. Same model: "minimax-h3", same mode, and the same model_params (prompt, aspect_ratio, duration, resolution, images, videos, audios) shown in the schema table below. The MCP server shares the same auth, model catalog, and local media handling as the CLI, so nothing about H3 itself changes — only how you invoke it.

Prerequisites

1. Install the AnyCap CLI (binary install is preferred over npm — fewer dependencies, faster startup):

   command -v anycap || curl -fsSL https://anycap.ai/install.sh | sh
   

2. Authenticate:

   anycap login
   # headless / SSH environments:
   anycap login --headless
   

3. Confirm video generation is enabled on your account:

   anycap status
   

   Look for "video.generate": true and "video.models": true in the feature list.

Step 1: Confirm the Model and Its Modes

MiniMax H3 supports three generation modes inside AnyCap. Check them before you write a prompt, since each mode accepts a different parameter set.

anycap video models minimax-h3

{
  "model": "minimax-h3",
  "display_name": "MiniMax H3",
  "description": "MiniMax H3 is a multimodal video generation model for creating short clips from text prompts or image, video, and audio references.",
  "operations": [{
    "operation": "generate",
    "modes": [
      { "mode": "text-to-video", "description": "Generate videos directly from text prompts." },
      { "mode": "image-to-video", "description": "Animate a reference image into a short video clip." },
      { "mode": "multi-modal-reference", "description": "Generate videos from image, video, or audio references with a guiding prompt." }
    ]
  }]
}

Step 2: Check the Parameter Schema for Each Mode

Always pull the schema before generating — parameters differ by mode.

anycap video models minimax-h3 schema --mode text-to-video
anycap video models minimax-h3 schema --mode image-to-video
anycap video models minimax-h3 schema --mode multi-modal-reference

Here's what each mode actually exposes today:

Step 3: Text-to-Video Generation

The simplest case — describe the shot in natural language and let H3 handle composition, motion, and native audio.

anycap video generate \
  --prompt "a neon-lit night market street in Taipei, rain reflecting on the pavement, ambient crowd chatter and sizzling street food sounds" \
  --model minimax-h3 \
  --mode text-to-video \
  --param aspect_ratio=16:9 \
  --param duration=8 \
  --param resolution=2k \
  -o night-market.mp4

Response shape:

{"status":"success","local_path":"/workspace/night-market.mp4","model":"minimax-h3","credits_used":12,"request_id":"req_abc123"}

That response shape is what you get every time. For proof it isn't just a JSON stub, here's an actual clip generated for this guide with the same command shape, a 5-second run of a coffee-dripper prompt, unedited model output:

Generated with anycap video generate --model minimax-h3 --mode text-to-video --param duration=5, 29 credits. Worth flagging: the prompt asked for ambient steam and ceramic sounds, and H3 generated atmospheric acoustic guitar music instead. Native audio generation here leans toward mood-appropriate soundtrack rather than literal sound-effect matching, so if a project needs specific Foley, test a few generations before committing to a duration.

Step 4: Image-to-Video Generation

Animate a still (a product photo, a poster, a UI mockup) into motion. Local files are auto-uploaded.

anycap video generate \
  --prompt "slow parallax push-in on the product shot, soft studio light shift, subtle ambient hum" \
  --model minimax-h3 \
  --mode image-to-video \
  --param images=./product-shot.png \
  --param duration=6 \
  --param generate_audio=true \
  --param aspect_ratio=1:1 \
  -o product-animation.mp4

Step 5: Multi-Modal Reference — H3's Signature Feature

This is what sets H3 apart from single-input models like Veo or Sora: you can combine a reference video (for camera movement), a reference image (for a character), and a reference audio track (for vocals) in one call, and describe the relationship between them in plain language.

anycap video generate \
  --prompt "Reference the camera movement from the first video, have the character in the image sing, with vocals matching the audio track" \
  --model minimax-h3 \
  --mode multi-modal-reference \
  --param videos='["./camera-ref.mp4"]' \
  --param images='["./character.png"]' \
  --param audios='["./vocal-track.mp3"]' \
  --param aspect_ratio=9:16 \
  --param duration=10 \
  -o composite-scene.mp4

This mode is the practical answer to "how do I keep a consistent character and camera style across clips without training a custom model." You just point H3 at prior assets and describe the relationship.

Where MiniMax H3 Fits Among AnyCap's Video Models

If you're choosing between models for a given AnyCap job, here's how H3 compares on the specs that matter for planning a shoot list:

Two takeaways: if native audio synced to the visual is a requirement, H3 is currently the only model in the AnyCap lineup that generates it as part of the video call rather than as a separate step. If you need resolution above 2K, Seedance 2.0 is the one to reach for instead.

Practical Use Cases

A few places this shows up in production, not just in demos:

• E-commerce teams turn a static product photo into a 6–10 second animated spot with generated ambient sound, ready to drop into a paid social placement — the shot below is a real example.
• Brand teams lean on H3's accurate text and logo rendering to produce on-brand title cards and animated posters without a separate compositing pass.
• Product and UI/UX designers animate a mockup to walk through a flow before a real prototype exists.
• Game studios generate short cinematic cutscenes or key-art loops straight from concept art.
• Anyone doing film-style opens can combine a reference clip's camera language with new subjects using multi-modal-reference mode.

Tips for Better Results

• Describe relationships, not just content. H3's captioning was trained on how context and target relate ("reference the movement from X, apply it to Y"), so that phrasing works better in multi-modal-reference mode than a flat scene description.
• Start short. Generate at 5–6 seconds first to check composition and audio before spending credits on a full 15-second clip; that's what the demo clip in this guide used.
• 2k is the only resolution option right now, and it's priced under a third of comparable mainstream 2K output, so there's no reason to default to a lower-fidelity model out of habit.
• Pull the schema before every new project. AnyCap updates model parameters over time, and anycap video models minimax-h3 schema --mode <mode> tells you exactly what's supported today rather than what was supported when this guide was written.

Frequently Asked Questions

Is MiniMax H3 free to use through AnyCap? Generation consumes AnyCap credits per the credits_used field returned with each request; check your plan for current credit costs.

Can I use a local file as an image or video reference? Yes. Pass a local path (e.g., --param images=./photo.png) and AnyCap auto-uploads it; a remote URL works the same way.

What's the difference between image-to-video and multi-modal-reference mode? image-to-video animates a single still image. multi-modal-reference accepts any combination of images, videos, and audio references simultaneously, letting you mix camera movement, subject appearance, and vocals from separate sources into one generated clip.

Does H3 generate audio automatically? In text-to-video mode, audio is generated natively as part of the clip. In image-to-video mode, use the generate_audio boolean parameter to control it explicitly.

Is MiniMax H3 open-weight? Yes — MiniMax has stated it is rolling out open model weights for H3, subject to applicable regulations, to support hardware compatibility and community customization. The hosted version is available immediately through AnyCap regardless of local weight availability.

Next Steps

Run anycap video models minimax-h3 schema --mode text-to-video right now and generate a 5-second test clip with a prompt describing your actual product or brand asset. Once you've validated composition and audio quality at the shorter duration, scale up to the full 15-second, 2K output for production use.