MiniMax H3 is the first major open-weight video model that generates native stereo audio in the same pass as the picture — not a TTS module bolted on after. Announced on July 31, 2026 and open-sourced on August 3, it’s a 33B-parameter dense transformer that takes text, images, video clips and audio tracks as input and outputs 4–15 second video at up to 2K resolution with 32 kHz stereo audio.
Where MiniMax M3 is the company’s LLM play, H3 is their video generation play — and the open-weight release is designed to do for video what Stable Diffusion did for images: give the community something real to build on.
Architecture
| Spec | Value |
|---|---|
| Model type | Dense single-stream Omni Transformer |
| Total parameters | 33B (~13B in AdaLN branches, cacheable) |
| Active inference parameters | ~20B |
| Text encoder | Qwen3-VL-32B (frozen) |
| Position encoding | 3D multimodal RoPE (time, height, width) |
| Precision | BF16 |
Output specs
| Value | |
|---|---|
| Duration | 4–15 seconds |
| Frame rate | 24 FPS |
| Resolution | 768p (open weights) / up to 2K (API via H3-Regenerate-2K) |
| Audio | 32 kHz native stereo |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and more |
| Languages | 11 stable (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish) |
Input modes
H3 ships two open checkpoints:
| Variant | Mode | What it accepts |
|---|---|---|
| H3-Base-FL2VA | First-and-last-frame | 0, 1 or 2 images → text-to-video, first-frame-to-video, first+last-frame-to-video |
| H3-Base-Ref2VA | Omni-reference | Up to 9 images, 3 video clips, 3 audio tracks (12 files max) |
What’s open vs what’s not
| Module | Status | Impact |
|---|---|---|
| H3-Base (FL2VA + Ref2VA) | Open (Hugging Face) | Core generation at 768p with native audio |
| H3-Context-IR | Closed (API-only) | Instruction processing — improves prompt interpretation |
| H3-Regenerate-2K | Closed (API-only) | Regenerates 768p output at 2K — the headline resolution |
| Sparse attention | Not in initial release | Native sparse attention mentioned in architecture but absent |
The honest framing: the base is open, but the pieces that produce the headline specs (2K resolution, optimized instruction following) are API-only. Local inference generates at 768p.
How it compares
- vs Seedance 2.0 / Seedance 2.5: Seedance leads the AA Video Arena on blind preference. H3’s advantage is open weights and omni-reference input (images + video + audio in one prompt). Seedance is proprietary.
- vs FLUX 3: FLUX 3 Video generates up to 20 seconds (vs H3’s 15) and BFL claims higher Elo scores. FLUX 3 Dev open weights are planned but unscheduled; H3’s base weights are available now.
- vs Veo 3.1: Veo has native audio at higher resolution with a public API at $0.40/sec. H3 offers open weights and a broader reference system at lower cost — but the full-quality comparison requires H3’s API tier.
- vs Hailuo 2.3: Both from the same company (MiniMax). Hailuo is the budget-focused API product; H3 is the next-generation model with open weights and omni-modal capabilities.
For the full video model landscape, see the video models directory.
Best for
- Self-hosted video generation with native audio for teams with GPU infrastructure
- Reference-driven video (character consistency, style transfer, voice cloning) from multimodal inputs
- Fine-tuning and customization on the open base weights
- Cost-effective API-based video production at 2K resolution
Pros & cons
Strengths
- First major open-weight video model with native stereo audio — not a pipeline bolt-on
- Omni-reference system accepts up to 9 images, 3 video clips and 3 audio tracks in one generation
- Single 33B dense transformer handles all modalities — no separate expert models per task
- 2K resolution by default via API at less than 1/3 the price of mainstream models
Limitations
- H3-Regenerate-2K (the module that lifts output to 2K) is not open-sourced — local inference caps at 768p
- H3-Context-IR (instruction-processing layer) also closed — API-only for full quality
- Community License is not MIT/Apache — commercial use capped at $20M annual revenue without separate agreement
- 15-second maximum duration — shorter than FLUX 3's 20s or Seedance 2.5's 30s
Frequently asked questions
When was MiniMax H3 released?
MiniMax announced H3 on July 31, 2026 as an API product, then published the base model weights to Hugging Face on August 3, 2026 under the MiniMax H3 Community License.
Is MiniMax H3 truly open source?
Open weights, not open source. The base model (H3-Base) ships under MiniMax's H3 Community License, which permits commercial use for organizations under $20M annual revenue. Two key modules — H3-Regenerate-2K (2K upscaling) and H3-Context-IR (instruction processing) — remain closed and API-only.
How does H3 relate to MiniMax M3?
Different product lines. MiniMax M3 is a 428B MoE large language model (text/image input → text output). H3 is a 33B dense video generation model (text/image/video/audio → video with native audio). They share a company but serve different purposes.
What hardware runs MiniMax H3 locally?
The full BF16 checkpoint is ~33B parameters. With AdaLN caching, active inference uses ~20B parameters. Community quantizations (e.g. MLX 6-bit at ~30GB on disk, ~16.5GB resident) exist for Apple Silicon. Expect to need at least one high-end GPU (80GB+ VRAM) for BF16 inference.
What resolution does the open model support?
768p natively. 2K resolution requires the H3-Regenerate-2K module, which is API-only and not open-sourced. The open weights generate at the shorter-side-768p default.