News / Model Watch · Video
MiniMax H3 Open Weights: Native A/V Video Model Lands for Local + API
MiniMax’s H3 is a general-purpose multimodal video generation model — not the company’s M3 text LLM. Official launch messaging (2026-07-31) + early-August open weights on Hugging Face / ComfyUI make this one of the most important open video drops of the summer.
Yes — write this as news. H3 unifies text / image / video / audio context and generates video with native stereo audio (not “video first, TTS later”). Hosted path: MiniMax / Hailuo APIs (often up to ~15s, up to 2K on hosted). Open path: MiniMaxAI/MiniMax-H3 + native ComfyUI workflows. Do not title it “open-source Kling 2.0/2.5” — that is not MiniMax’s product name.
H3 vs M3 (stop the name mix-up)
| Name | Modality | Role |
|---|---|---|
| MiniMax H3 | Video + native audio (omni-modal) | This brief. Generation model; open weights + API |
| MiniMax M3 | Text LLM | Separate product line on MiniMax’s model menu — not a video SKU |
Capability highlights (official + ecosystem)
Primary: MiniMax H3 launch post. Secondary: HF card, ComfyUI, X creator demos — not invented benchmarks.
Native stereo audio-video
Joint generation of picture + stereo sound (dialogue / SFX / music framing) in one pass — a major differentiator vs pure silent video models.
~4–15s · up to 2K (hosted)
Official: up to about 15 seconds at up to 2K on hosted paths. Open-weight local runs are often discussed at lower default res (e.g. ~768p) unless using upscale/regenerate paths.
Multimodal references
Text + images + video + audio as context. Community/model notes often cite multi-asset reference (e.g. several images + video + audio clips) for controlled generation/editing.
Open weights + ComfyUI
Weights on Hugging Face; ComfyUI native nodes/workflows within days. Community already ships Turbo LoRAs and quant packs — classic open-source velocity.
“Open Kling” vs “open Seedance” — what X actually says
- Official framing: open-weight multimodal video model from MiniMax / Hailuo — not licensed or named as Kling.
- Community quality talk: frequent side-by-sides with Seedance 2.0 / 2.5 (“close quality,” “open Seedance,” arena Elo chatter). Treat as anecdotes.
- Kling comparisons: appear as casual creator chat, not a MiniMax product claim. Do not publish “开源版可灵” as fact.
- Price narrative: MiniMax claims strong price-performance vs “mainstream” closed video models at 2K/768p — verify in your console; we do not republish unofficial $/s as ours.
Access map (as of 2026-08-08)
| Surface | Status | Notes |
|---|---|---|
| MiniMax blog / Hailuo product | Live | Official H3 post |
| Hosted API | Live | platform.minimax.io / CN console video generation docs |
| Hugging Face weights | Live | MiniMaxAI/MiniMax-H3 · Community License (read territory/commercial clauses) |
| ComfyUI | Native support | Official workflows + community quant/Turbo LoRA |
| ChinaModelAPI relay | Pending health-check | No invented model ID — join waitlist |
Builder checklist
- Decide path: hosted API (2K convenience) vs local weights (VRAM + license).
- Read the Community License carefully — commercial thresholds and regional exclusions appear in community repacks; trust the official LICENSE file.
- Budget VRAM realistically: full BF16 stacks discussed in the wild often exceed a single consumer GPU without quant/pruning.
- Compare creatively against Seedance 2.5 for closed flagship quality and against cheaper 2.0 Mini for iteration cost.
- If you need OpenAI-compatible Chinese video routing later, stay on the ChinaModelAPI waitlist — we only publish IDs after routing works.
FAQ
H3 or M3?
H3 = video. M3 = MiniMax text LLM. Different product lines.
Open Kling?
No. Community may compare quality; product name is MiniMax H3 / Hailuo line.
Fully open-source everything?
Base weights opened; some hosted pieces (e.g. higher-res regenerate / context services) may remain API-only — check the model card.
Related on this site
Primary sources
- MiniMax — MiniMax H3 launch post (2026-07-31)
- Hugging Face — MiniMaxAI/MiniMax-H3
- X: @MiniMax_AI / @Hailuo_AI / @ComfyUI August 2026 open-weights + Turbo LoRA threads