What Wan is
Wan is a video generation model that has performed strongly on blind preference benchmarks and is also released with open weights. That combination — competitive quality plus the option to run it yourself — is what distinguishes it.
- Text to video and image to video.
- Strong performance on independent blind preference testing.
- Open weights available, so it can be self-hosted.
- Suitable as a base for fine-tuning and adapters.
Hosted versus self-hosted
| Route | Choose when | Trade-off |
|---|---|---|
| Hosted | convenience and elasticity matter | You accept vendor and pricing risk |
| Self-hosted | Control, reproducibility or volume economics matter | You take on hardware and maintenance |
The control argument: if you need to prove output has not changed, or your footage is confidential, a model you run yourself has advantages no hosted service can match.
Prompt approach
- Subject, motion, camera, lighting, duration — structured and short.
- Name one camera move or say the camera is static.
- Keep clips short for stability.
- Generate variants and select.
Structured shot
A glass teapot on a dark slate table, steam rising in a thin column, completely static camera, single warm light from the right, deep shadow, minimal and calm, 5 seconds.
Hardware considerations
- VRAM is the binding constraint, determined by model size plus your target resolution and clip length.
- Quantisation reduces memory requirements at some quality cost, and often decides whether a model runs at all on available hardware.
- Measure end-to-end time on your own hardware — published benchmarks rarely match real setups.
- Generation speed matters for workflow; a very slow model is unusable regardless of quality.
When self-hosting is worth it
-
Measure your real volume
Clips per week and minutes of output.
-
Price both routes on standard rates
Include GPU amortisation, power and your maintenance time for self-hosting.
-
Weight reproducibility
If you need to prove output is unchanged, this may justify a premium.
-
Check the licence
Confirm commercial use is permitted for your situation.
-
Prototype for a week
Rent hardware and run your real workload. Measured numbers settle the argument.
Practical limits
- Self-hosting is usually more expensive at low volume because of idle capacity.
- You do not receive the invisible improvements hosted providers ship.
- The very largest models may not be practical to run locally.
- Benchmark position does not guarantee performance on your specific prompts.