MiniMax H3

About

MiniMax H3 is an omni-modal AI video model that generates video and synchronized stereo audio together. Sogni's FL2VA workflows support text-to-video, first-frame image-to-video, and first-and-last-frame video generation.

Sogni runs ComfyUI's pruned FP8-scaled diffusion model with the NVFP4 Qwen3-VL 32B encoder. The initial release is routed to 32 GB NVIDIA workers while Sogni validates the newer DynamicVRAM runtime across the full GPU fleet. Output is fixed at 24 fps and 32 kHz stereo audio, approximately 5.17–15.08 seconds, with 1344×768 and 768×1344 as the primary landscape and portrait sizes.

Sogni has received written authorization directly from MiniMax to offer MiniMax H3 through the Sogni platform. Built with MiniMax H3, the integration retains MiniMax attribution and links to the official model card and license.

This release runs the open H3-Base FL2VA weights locally. MiniMax's Ref2VA checkpoint, hosted H3-Context-IR service, and H3-Regenerate-2K stage are outside this release, so local Sogni output is 768p-class rather than native 2K.

Variants

MiniMax H3 ships as 3 workflow variants. Pick the input mode that fits, or start with the default.

Prompting tips

Describe one continuous shot with enough information for both pictures and sound: subject, action, camera, lighting, ambience, effects, and dialogue.

• Action — Give the shot one coherent motion thread and a clear final beat.
• Camera — Choose one deliberate move such as a slow push-in, orbit, pan, or locked-off composition.
• Audio — Describe dialogue, voice character, music, ambience, and synchronized sound effects in the same prompt.
• Dialogue — Put spoken lines in quotation marks, identify the speaker, and keep the word count realistic for a 5–15 second clip.
• First + last frame — Describe the movement that connects the two supplied compositions and the visual beat that should land on the final frame.

Built with MiniMax H3 — authorized for Sogni

Sogni has received written authorization directly from MiniMax to offer MiniMax H3 through the Sogni platform. The integration credits MiniMax, links the official MiniMax H3 model card, and retains the MiniMax H3 Community License Agreement.

How MiniMax H3 generates video and audio together

H3-Base packs text, visual, and audio representations into one multimodal sequence. A dense 33-billion-parameter single-stream Omni Transformer jointly predicts video and audio latents; separate visual and audio VAEs then decode the result into moving pictures and 32 kHz stereo sound. The text path uses Qwen3-VL-32B hidden states, while multimodal rotary position embeddings preserve temporal and spatial relationships.

The FL2VA checkpoint accepts zero images for text-to-video, one image for first-frame animation, or two images for first-and-last-frame control. Sogni exposes those as three explicit workflows so asset requirements, pricing, and worker routing remain predictable.

Consumer-GPU optimization in ComfyUI

ComfyUI prunes modulation weights that can be replaced by a lookup table, publishes FP8-scaled and INT8 quantizations, and combines custom kernels with dynamic VRAM offloading. ComfyUI reports that its smallest model set reduces the total memory footprint from 123.6 GB at full precision to 42.5 GB. Sogni pairs the pruned FP8-scaled diffusion model with the NVFP4 AWQ text encoder and initially routes jobs to 32 GB NVIDIA workers while the newer DynamicVRAM runtime is validated across every supported GPU and Windows/WSL2 configuration.

768p local output versus the hosted 2K pipeline

The open H3-Base model produces 768p-class video. MiniMax's full 2K system adds the hosted H3-Context-IR preprocessing service and H3-Regenerate-2K, which regenerates the base result with the original context instead of applying a conventional super-resolution pass. Those two hosted modules are not part of the open local release, so Sogni does not label this workflow as native 2K.

Read the ComfyUI MiniMax H3 day-zero article for the local optimization details, or open the Sogni MiniMax H3 integration guide for exact model IDs, frame math, GPU requirements, and prompting guidance.

Pricing

MiniMax H3 is included on every Sogni plan — generate credit-free under fair use, alongside 200+ other models. Pay-as-you-go Spark packs are also available for eligible Spark renders.

API

One Sogni API key reaches every model on the Supernet — call MiniMax H3 with the exact model id for the variant you want.

import { SogniClient } from '@sogni-ai/sogni-client';

const client = await SogniClient.createInstance({
  appId: crypto.randomUUID(),
  apiKey: process.env.SOGNI_API_KEY,
  network: 'fast',
});

const project = await client.projects.create({
  type: 'video',
  modelId: 'minimax-h3-fl2va-fp8_t2v',
  positivePrompt: 'a slothicorn surfing a wave of liquid paint, slow push-in, cinematic',
  numberOfMedia: 1,
  duration: 5,
});

const [url] = await project.waitForCompletion();
console.log(url); // result link — download within 24h

curl https://api.sogni.ai/v1/creative-agent/workflows \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SOGNI_API_KEY" \
  -d '{
    "input": {
      "title": "First render",
      "steps": [{
        "id": "step1",
        "toolName": "generate_video",
        "arguments": { "prompt": "a slothicorn surfing a wave of liquid paint, slow push-in, cinematic", "videoModel": "minimax-h3", "duration": 5 }
      }]
    },
    "confirm_cost": true
  }'

Example uses minimax-h3-fl2va-fp8_t2v. Use any exact model id from the Variants table above. Full reference at docs.sogni.ai.