MiniMax-H3 (Hailuo 3.0): A 33B Omni-Modal Video Model With Native Stereo Audio

Key points

What MiniMax-H3 actually is

• Release and provenance: Launched 2026-07-31 by MiniMax (Shanghai) under its "Hailuo" video line (Hailuo 01 → 02 → H3 / Hailuo 3.0). Open weights shipped 2026-08-03 on Hugging Face (MiniMaxAI/MiniMax-H3) and ModelScope.
• Core network: A 33B-parameter dense, single-stream H3-Omni-Transformer (50 layers, hidden dim 5376). About 13B parameters live on AdaLN branches and can be cached at inference.
• Inputs: Any combination up to 9 images + 3 videos + 3 audio clips, prompt limit 7,000 characters, 11-language dialogue.
• Outputs: Up to 15s, 24fps, max 2K; native stereo audio at 32kHz. Default open weights output 768p; 2K requires the closed H3-Regenerate-2K.
• Text encoder: Qwen3-VL-32B (layer-50 hidden states, nvfp4_awq), genuinely Apache 2.0 open source.
• Three components: H3-Context-IR (instruction understanding, not open), H3-Base FL2VA/Ref2VA (open), H3-Regenerate-2K (2K upscaler, not open).

Key fact corrections vs. prior reporting

• Not a Mamba/SSM hybrid. Earlier posts that called H3 an SSM+attention hybrid conflated it with MiniMax-01 (which uses lightning/linear attention). H3 is a plain dense Transformer unified through 3D multimodal RoPE.
• No local 2K. Open H3-Base outputs 768p; 2K is cloud-only via the closed H3-Regenerate-2K.
• Stack memory is much larger than 37 GiB. "33B" refers to the DiT body. The full BF16 stack is roughly 144GB (DiT ~66.3GB + encoders ~66.7GB + video VAE ~10.4GB + audio VAE ~0.6GB). A quantized ComfyUI distribution is ~42.5GB.
• Not OSI open source. Weights are downloadable under a Community License; the context-understanding and 2K modules are not released.
• Local hardware claims are softer than marketed. RTX 3060 12GB with offloading can produce output (5s 480p in ~5 minutes) but is heavily quantized. Apple Silicon MLX builds download ~115GB; ≥192GB unified memory is recommended, 512GB verified.

Architecture overview

• Single-stream omni-transformer: text, image, video, and audio tokens share one backbone; no modality-specific attention or FFN.
• Visual VAE (f16t4d24): causal in time, 16× spatial and 4× temporal compression, 24 latent channels; a 1×2×2 patchify yields an effective 32× spatial downsample before the Transformer.
• Audio VAE: separate L/R encoder-decoders merged into stereo; 32kHz compressed to ~40Hz latent token streams.
• AdaLN cache: ~13B parameters on the AdaLN modulation path can be precomputed, letting MLX ports drop resident memory from ~134GB to ~40GB.
• Sparse-attention roadmap: MiniMax has discussed sparse attention for follow-ups, but the current open runtime uses full attention; do not directly compare local full-attention latency with the cloud's internal kernels + sparse path.

Capabilities and benchmarks

• Differentiator: native stereo dialogue, SFX, and music generated jointly with video in one forward pass — rare among open video models (HunyuanVideo 1.5 is silent).
• Ref2VA / R2V mode: combines up to 9 images + 3 videos + 3 audio clips with natural-language relationship descriptions.
• Instruction editing via H3-Context-IR works well but is not open; it requires the API or a self-built prompting harness.
• Audio caveat: independent tests (e.g., Simon Willison on M5 Max) note that without explicit sound direction in the prompt, audio degrades to "speech-shaped noise." Prompt engineering for dialogue + ambience + score is essential.
• Benchmarks (directionally, not absolute): Video Arena ranks H3 #1 among open video models (~1476 in image-to-video, 2 points behind Seedance 2.0); Artificial Analysis places H3 ~Elo 1130 in audio-video editing. Treat Elo-style scores as prompt-set-specific evidence, not ground truth.

Licensing — read this carefully

H3 weights are not OSI open source. The Community License (effective 2026-08-02) imposes:
• Geographic exclusions: Weights are licensed globally except the US, EU, UK, and South Korea, which require separate authorization. Hosted APIs remain globally available.
• Revenue threshold: Annual revenue > $20M requires prior written authorization.
• UI attribution: Commercial UIs must prominently display "MiniMax H3".
• No model laundering: Outputs cannot be used to train, fine-tune, distill, or improve other AI models.
• Distribution: Original license, NOTICE, and modification logs must accompany any redistribution; governed by Hong Kong law.
• The four excluded regions reflect ongoing generative-video copyright litigation with Disney, Universal, and Warner — not a geopolitical posture, per developer-relations lead Ryan Lee.
• Outlook: MiniMax announced on 2026-08-09 that it plans to migrate to Apache 2.0 once copyright issues resolve, and intends to open H3-Regenerate-2K, 4/8-NFE low-step variants, and a text-to-image model.

Ecosystem and deployment

• Day-0 tooling: ComfyUI (Comfy-Org/MiniMax-H3 with T2V / FLF2V / R2V workflows and quantized weights), Diffusers pipelines, SGLang deployment guide, and h3.c (antirez's pure C + Metal port for Apple Silicon, MIT).
• Hardware brackets:
• Flagship desktop (RTX 4090/3090 24GB + quantization): high-resolution runs, best quality.
• Consumer (RTX 4070 Ti / 3060 12GB + int8 + offload): 5s 480p in 1–5 minutes.
• Edge / mini (DGX Spark FP8): ~89GB load, FL2VA ~155s in tests.
• Apple Silicon (MLX): ≥192GB unified memory recommended, 512GB verified.
• Server: 8×141GB-class nodes for full BF16 stack.
• Local minimum: NVMe SSD, 32GB system RAM (64GB+ recommended). Remember: local 2K does not exist — open weights cap at 768p.
• API path: MiniMax API globally, plus fal, Segmind, and Chinese MaaS providers (e.g., Qiniu, model id minimax/minimax-h3, billed per second, around ¥0.8/s for 2K).

Competitive landscape

| Model | Resolution | Native audio | Open weights | Local run | Notes | |---|---|---|---|---|---| | MiniMax H3 | 768p local / 2K cloud | ✅ stereo 32k | ✅ (Community Lic.) | ✅ (quantized) | Strongest multi-reference control | | Sora 2 / Pro | up to 1080p | ✅ | ❌ | ❌ | Closed; $0.30–0.50/s | | Veo 3.1 | up to 4K | ✅ | ❌ | ❌ | Closed; $0.40–0.75/s incl. audio | | Seedance 2.0/2.5 | up to 1080p | ✅ | ❌ | ❌ | Closed; ~3× H3 price | | HunyuanVideo 1.5 | 480/720p +1080p SR | ❌ | ✅ | ✅ | Local silent-video leader | | Wan 2.7 | — | ✅ | ✅ | ✅ | Top-3 on AA | | LTX-2.3 | — | ✅ | ✅ | ✅ | Joint audio/video, trainable | | Kling / Runway | 1080p | ❌ | ❌ | ❌ | Closed |

H3's edge is the combination: open weights + local control + native audio + multi-reference control + ~1/3 the price of closed competitors — not a single-axis quality win.

Honest limits

1. "Open" means weight-accessible under a custom Community License; two core modules remain closed and four jurisdictions are excluded. 2. Local generation tops out at 768p. 3. Audio quality scales with prompt effort — without explicit dialogue/SFX/score guidance, outputs regress to noise. 4. Marketing softens hardware reality: 12GB works but is slow; Apple needs 192GB+; full BF16 stack is ~144GB. 5. Elo-style benchmarks are prompt-pool specific, not absolute. 6. The open runtime is full-attention; cloud latency benefits from internal sparse kernels and a two-stage pipeline.

Verdict

MiniMax-H3 is the first open-weights video model to productize the hardest problem in generative media — synchronized audio and video from a single forward pass. Its strategic weight lies in reopening a frontier previously held closed by Sora and Veo, riding the 2026 Chinese open-weights wave alongside DeepSeek, Qwen, and Kimi. The realistic shortlist criteria are local-first, data-on-device, long-term-controllable workflows. Anyone expecting "download → 2K → 12GB smooth → global unrestricted" should temper expectations — H3 is a spearhead of the open-weights wave, not an unconstrained free lunch.

Sources

• MiniMax official open-source blog (2026-08): H3-Omni-Transformer, VAE, AdaLN, component split.
• MiniMax official AMA / Reddit (2026-08-09): Apache 2.0 migration intent, Regenerate-2K preview, sparse-attention roadmap.
• Hugging Face MiniMaxAI/MiniMax-H3 model card and LICENSE (Community License, effective 2026-08-02).
• metirai.com, nxcode.io, domoai.app, atlascloud.ai (2026-08): architecture, dual-benchmark #1 ranking, license breakdown, system limits.
• pandaily.com, newshunt.io (2026-08): pricing vs Seedance 2.0, dual-RTX-5090 local runs, API pricing.
• dev.to, ai-indeed.com (2026-08): ComfyUI deployment, hardware brackets, ~144GB full-stack memory.
• Qiniu MaaS tutorial (news.qiniu.com): ComfyUI dual paths, API pricing (¥0.8/s for 2K).
• Tencent News (2026-08-04, 2026-08-10): Video Arena ranking, Ryan Lee's response on the four excluded regions.
• Earlier h3.c deep dive (2026-08-11): antirez's C/Metal engine; architecture description corrected in this article.