QWE AI Academy Text to video AI offline, on your box: OpenVideo v0.1.0 + MiniMax H3. End state is simple – an MP4 from open-video run, no per-second API meter.
OpenVideo is a thin “Ollama for video” CLI over ComfyUI: install → pull → status → run. Not a hosted studio. Skip the SaaS roundup; this guide is deployment only.
System requirements before you touch git
As of the v0.1.0 docs (changelog tag dated 2026-09-22), plan for a real workstation – not a thin laptop.
| Resource | Minimum that can limp | What you actually want |
|---|---|---|
| OS | Linux or Windows WSL2 | Native Linux + recent NVIDIA driver |
| GPU | ~12 GiB VRAM with low-VRAM offload (slower, fiddly) | ≥22 GiB for default INT8 ConvRot pull |
| Disk | ~54 GB for H3 weights alone (install script) | 60 GB+ free NVMe (site quickstart ballpark) plus room for outputs/cache |
| RAM | Not hard-coded in the OpenVideo README | Enough headroom for ComfyUI offload when VRAM is tight – size to your other apps |
| Network | Resumable weight download | Stable link; first pull is multi-tens of GB |
VRAM table in the OpenVideo README: pull only ships checksum-verified INT8 ConvRot. The 12-22 GiB band leans on automatic ComfyUI low-VRAM offload. W4 (~9-12 GB class) and NF4 (<9 GB) stay manual experiments – recommend-quant may name them; the installer never fetches them.
macOS: clone and dry-run are fine. Full H3 generation stays NVIDIA/Linux-first. Native Windows generation is unverified – WSL2 is the supported path.
Official source (pin the tag)
Do not treat the website one-liner as truth for this exact release. The project notes the site installer updates on its own schedule and can lag the GitHub tag.
# Preferred: pin v0.1.0
git clone --depth 1 --branch v0.1.0 https://github.com/open-video-ai/open-video
cd open-video
bash scripts/install.shThat script wires ComfyUI and starts the resumable H3 weight fetch (~54 GB per install notes; site quickstart still ballparks ~60 GB disk). If you accept version drift: curl -fsSL https://open-video.ai/install | bash from open-video.ai.
Manual path without the shell installer:
git clone --depth 1 --branch v0.1.0 https://github.com/open-video-ai/open-video
cd open-video
pip install -e .
open-video pull h3First-time configuration (bare-minimum setup)
Three checks after install: weights present, engine reachable, license accepted for your geography.
open-video recommend-quant– confirm tier guidanceopen-video pull h3– finish or re-verify weights (v0.1.0 manifest: sizes + checksums; re-run if a prior download died mid-way)- Non-default engine:
export OPEN_VIDEO_COMFYUI=http://127.0.0.1:8188 - Image / first-last frame jobs: set
OPEN_VIDEO_COMFYUI_INPUTto a directory owned by that ComfyUI instance so concurrent runs do not clobber shared frame names
Pro tip: Read docs/WEIGHTS_LICENSE.md before any commercial render. OpenVideo code is Apache-2.0. H3 weights are not. The MiniMax H3 Community License – as quoted there from upstream text fetched 2026-08-07 – limits the default grant to the Applicable Territory (worldwide excluding the EU, UK, Republic of Korea, and United States) and requires prior written authorization above US$20M yearly revenue on commercial products/services. Confirm live terms on the upstream MiniMax H3 card. This is not legal advice.
Think of the weight license like a venue lease: the sound system (CLI) moves freely; the band (H3 checkpoint) only plays where the contract says.
Verify the text to video AI install works
open-video status # or: open-video ps
open-video run "sunset waves" --dry-run
open-video run "A lone astronaut planting a flag on a red dune at dusk" --duration 8
# optional explicit output
open-video run "waves at sunset, golden hour" --duration 10 --model h3 --output out.mp4status/ps: engine health + weight inventory. Dry-run plans and validates without GPU spend. A good run writes an MP4. Local adapter default still lands ComfyUI around http://127.0.0.1:8188.
H3 open path targets short clips with native audio (community/system notes often cite ~4-15 s; default short side on Base is about ~768p). Structured multi-field prompts beat vibes-only one-liners here – that’s the control that actually moves clip quality on a local stack, not a longer SaaS credit pack. Opt-in VLM judge: set OPEN_VIDEO_VLM_URL / model / key; without them the verdict is SKIPPED (score 0), not a fake PASS. OPEN_VIDEO_JUDGE_RETRIES buys extra takes when the judge is live.
Common install errors and fixes
- Pull finished but hash/size complaints – Re-run
open-video pull h3. v0.1.0 checks the packaged manifest (sizes + SHA-256), including same-size leftovers from a broken mirror. - CUDA OOM mid-sample – Cut resolution/duration first. On 12-22 GiB stay on INT8 + low-VRAM offload; close other GPU apps; restart ComfyUI to clear cached allocations. Sub-9-12 GB W4/NF4 paths are experimental and outside the default pull.
- Engine “hangs” with no error – Read the ComfyUI console, confirm
OPEN_VIDEO_COMFYUI, and hit the port directly. - I2V/FL frames overwrite each other – Point
OPEN_VIDEO_COMFYUI_INPUTat the correct server-owned input dir. - Judge always “passes” – No VLM URL/model/key → honest
SKIPPED/ score 0. Intentional. - Windows native path fails – Same clone, WSL2. Native execution remains unverified.
Upgrade and uninstall
Upgrade: newer published tag → reinstall editable package if you used pip → open-video pull h3 again so manifests match. v0.1.0 has no fancy migrator; treat weights as replaceable artifacts.
Cleanup: stop ComfyUI; delete the clone; wipe installer/HF cache dirs you used; pip uninstall open-video if applicable; reclaim the ~54 GB+ weight footprint and leftover MP4s.
Will multi-model backends show up before long-form stitching feels production-ready? Only H3 pull/run ships in v0.1.0 – measure your own GPU hours before betting a pipeline on scaffolding.
Next: pin v0.1.0, bash scripts/install.sh, then open-video status and a dry-run before the first eight-second run.
FAQ
Is OpenVideo the same as MiniMax Hailuo cloud?
No. Local CLI over ComfyUI with open H3 weights. Hailuo/API is separate hosting, billing, and feature set – including pieces that are not fully open (for example some 2K regenerate flows).
How much VRAM do I need for a first successful clip?
Want the path pull actually installs? ≥22 GiB for default INT8 ConvRot. A 24 GB card is the boring “it just worked” zone. On ~16 GB, INT8 + automatic low-VRAM offload is the documented band – expect longer wall time and OOM if anything else holds VRAM. W4/NF4 remain manual experiments.
Can I use outputs commercially after a local install?
Apache-2.0 covers OpenVideo’s code, not the checkpoint. Geography, attribution, and revenue hooks live in the MiniMax H3 Community License summary under docs/WEIGHTS_LICENSE.md (see the pro tip above). Re-read the live upstream LICENSE on the model card before client work – terms can change; a tutorial is not a grant.