open-video
library Model backends

Every open video model.
One director.

open-video is model-agnostic: one autonomous core (plan → craft → validate → judge → refine → stitch) driving any pluggable backend. H3 ships today; Wan 2.2 and LTX-2.3 drop in without touching the brain.

3 backends · 1 active · Apache 2.0 platform · local-first

MiniMax H3 / h3

Active

The #1 open video model — Artificial Analysis Arena T2V #2 / I2V #3 overall, #1 open. Native stereo audio, top-tier prompt adherence, every mode (T2V / I2V / FL2VA / R2V). The baseline backend open-video ships against today.

Try H3  → free · no sign-up
$ open-video "a neon koi swimming through rain, slow dolly" --model h3
Key specs
Params33B
Max resolution768p short edge
AudioStereo 32kHz native
FPS24 17k+5 VAE
Max duration15s per shot
LicenseOpen weights
native stereo audio prompt adherence Arena #1 open all modes (T2V/I2V/FL2VA/R2V)
Community gallery · generated on H3 Browse all →
neon koi · rain dolly
lunar rover · 21:9
jazz club · w/ audio
paper crane · macro

Wan 2.2 / wan2

Coming soon

The Apache-2.0 clean-license anchor — a fully permissive backend with no license ambiguity, strong physical realism, and a clean fit for commercial pipelines. The first alternative backend to prove open-video's model-agnostic seam.

Notify me on roadmap
$ open-video "..." --model wan2 # backend in backends/wan2/
Target specs
ParamsTBD
Max resolutionTBD
AudioLimited vs H3
FPSTBD
TierQuality physics
LicenseApache 2.0
physical realism Apache 2.0 clean license commercial-friendly
Community gallery · coming when the backend lands Track progress →
water sim · physics
cloth & fabric
product shot
nature b-roll

LTX-2.3 / ltx

Coming soon

The real-time speed tier — fast iterations, low latency, cheap drafts. Not the quality ceiling of H3, but the right tool for storyboards, previews, and the inner loop of the judge→refine cycle where you generate many takes.

Notify me on roadmap
$ open-video "..." --model ltx # speed tier
Target specs
ParamsTBD
Max resolutionTBD
AudioNone tier focus
FPSTBD
TierReal-time speed
LicenseOpen weights
real-time speed low latency cheap drafts & storyboards
Community gallery · coming when the backend lands Track progress →
storyboard frame
rapid preview
best-of-N drafts
animatic pass
seam Model-agnostic by design

The brain never changes.
Only the backend does.

open-video's core — planner, crafter, validator, judge→refine loop, stitcher — is model-agnostic. Each model is a plugin under backends/<model>/ that declares its capabilities, prompt grammar, and constraints. H3 today, Wan 2.2 and LTX-2.3 tomorrow, your fork the day after.

backends/h3/capabilities · prompt_grammar · workflows · constraints # active
backends/wan2/Apache-2.0 quality backend # roadmap
backends/ltx/real-time speed backend # roadmap
core/planner · crafter · validator · judge · stitcher # shared, untouched
Add a backend  → Read the docs