what sign up
nixm nixm

← learn

open-source ai video — the real cost of running it yourself

August 1, 2026

The weights are free; the hardware isn't. Wan 2.2's flagship wants roughly 54GB of VRAM at full precision, and quantizing down to a consumer card trades quality and time. If you want the clip and not the pipeline, nixm renders HD with sound for one non-expiring credit.

Free weights, priced hardware

Open-weights video models are genuinely free to download and, in Wan's case, permissively licensed for commercial use. The bill just moves. Wan 2.2's 14B flagship needs somewhere around 54GB of VRAM at full precision — datacenter territory, above any consumer card. Quantization plus offloading the text encoder to system RAM brings it down to roughly 6 to 8GB at 480p, or 12 to 16GB at 720p, which is where most self-hosters actually run.

The cost is minutes, not just gigabytes

On a 16GB card, a 720p clip lands in something like two to four minutes — respectable, and that's with the model already downloaded, the workflow already wired, and the settings already right. Get one wrong and you wait the same minutes for unusable output. Hardware without CUDA fares worse: one documented run on a 64GB M1 Max took 82 minutes to produce two seconds of video. Iteration speed, not peak quality, is what self-hosting usually costs you.

The landscape moves under you

LTX-2.3 is the fast option and the first open model generating synced audio in one pass, though its official baseline starts around 32GB before quantization. Hunyuan handles faces well — see the Hunyuan breakdown. Older names have already aged out: Mochi 1 broke ground and has been surpassed, and Stable Video Diffusion was pulled from Stability's API in 2025. Running local means tracking all of this yourself.

When self-hosting is right — and when it isn't

If you're rendering constantly, own the GPU, and enjoy the workflow graph, local wins outright and nothing here argues otherwise. If you want a striking clip on a Tuesday evening, the setup is the product you didn't want to buy. nixm renders HD with native sound in about two minutes for one credit that never expires — no card, no quantization, no node graph. The comparison in plain numbers: how AI video credits work.

Try it without installing anything

Direct a scene in the cinematic studio — first HD video with sound is free.

frequently asked

Can I run AI video generation on my own PC?

Yes, with enough VRAM. Quantized builds of Wan 2.2 run from roughly 6 to 8GB at 480p, or 12 to 16GB at 720p.

How much VRAM does Wan 2.2 need?

Around 54GB at full precision for the 14B model. Quantization plus CPU offload of the text encoder brings it into consumer range.

How long does local AI video take?

Roughly two to four minutes for a 720p clip on a well-configured 16GB card. Non-CUDA hardware can be dramatically slower.

Is open-source AI video actually free?

The weights are. The GPU, the electricity, the setup time and the failed renders are the real cost.

Which open model is best right now?

Wan 2.2 leads on silent-video quality, LTX-2.3 on speed and native audio. Both move fast enough that any answer dates quickly.

Can I try it free?

Yes — nixm includes one free HD video with sound on a new account, no card required.

try it now — nixm makes cinematic ai video, with sound. no subscription. first video free.

open the studio →

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or
+