Wan 2.7 hosted in Versely versus self-hosting

Wan 2.7 the open weights and Wan 2.7 the catalog row are the same family. They are not the same ops decision. Hosted credits buy you no GPU and a bill. Local ComfyUI buys you VRAM, a machine that stays on, and no per-clip invoice. Break-even is a volume question. Most brand teams should stay hosted.

This page is that fork. It is not a quality bake-off against Veo or Kling — that comparison already exists as Wan 2.7 versus closed-source models. Quality is not what changes when you move the same family from a hosted endpoint to a tower under a desk.

What you are actually buying

Hosted. You pick Wan in Versely, pay credits, get a clip. No driver dance, no VRAM math, no one on the team who "knows Comfy." The bill is per generation. Idle cost is zero. When the model is the wrong job you route to a different row on the same bill instead of standing up a second stack.

Self-hosted. You download the weights, run them in ComfyUI or an equivalent, and generate on hardware you own or rent by the hour. There is no per-clip invoice from a video API. There is a GPU, electricity, cooling, queueing, and a person who keeps the node alive. The Apache 2.0 weights make that legally possible. They do not make it free.

The family is the same in both cases: text-to-video, image-to-video, reference-to-video, instruction edits. Self-hosting does not unlock a secret Wan. It changes who owns the machine.

Break-even is a volume question

Hosted looks expensive by the clip and cheap by the week if you generate in bursts. Self-hosted looks free by the clip and expensive by the month if the card sits idle or the operator's time is real.

Run the arithmetic on your actual cadence, not on a theoretical 24-hour queue.

• A few dozen clips a week, mixed models, mixed jobs. Hosted. The Wan row is one of several you will touch. Standing up a node for a minority of the slate is theatre.
• A daily Wan-shaped pipeline, one look, one operator who already runs Comfy. Self-host starts to make sense, and only on the Wan-shaped part of the slate. Everything that is not Wan still needs a hosted path.
• Burst campaign, then quiet. Hosted. You do not want a GPU as a retainer for a week of chaos and three weeks of nothing.
• Fine-tune on your footage. That is a self-host (or a private cluster) conversation, because hosted inference does not give you a training loop. Fine-tune is a different product from "we generate a lot."

Do not convert credits to dollars and compare them to the street price of a used 4090. The 4090 is not the cost. The cost is the 4090 plus the hours to install, the hours to recover a failed run, the queue when two people render at once, and the clips you did not make because the node was down. Brand teams consistently undercount that labour. That is why the hosted bill, which looks like the expensive option on a spreadsheet, is usually the cheaper operating system.

Most brand teams should stay hosted

A brand team is not a lab. The constraints are calendar, brand kit, and "someone else has to rerun this next month." Hosted Wan satisfies those. The clip is reproducible from a project, not from an unnamed custom node a contractor left behind.

Stay hosted if any of the following is true:

• More than one person has to generate, and they do not share a Comfy graph.
• You already fail over between Wan, Veo, Seedance, Kling, Flux. A single local checkpoint cannot do that.
• Legal wants a vendor and a DPA, not a tower.
• Nobody on staff wants to be on-call for CUDA.

Self-host if all of the following are true: the slate is Wan-shaped, volume is high and steady, someone already operates the box, and the alternative is a per-clip bill that you have measured rather than feared.

"We might generate a lot someday" is not volume. It is a wish. Hosted until the wish is a calendar.

What self-hosting does not buy you

It does not buy better Wan. Same weights, same failure modes, same prompt obedience. If hosted Wan misses a hand, local Wan misses the hand.

It does not buy a studio. Captions, publishing, avatars, first-last product turns, always-on dialogue — those are other models. A local Wan box is a Wan box.

It does not, by itself, satisfy "the file cannot leave the machine." Hosted Wan sends the prompt and the references to a provider. Local Wan keeps them on the box if you actually air-gap it. A Comfy install that still pulls from a random Hugging Face mirror and logs to a cloud sync folder is not an air gap. If the brief is "the plate cannot leave the building," that is a different constraint — Hunyuan 1.5 when the file cannot leave the machine — and Wan-on-Versely cannot meet it. Most briefs are not that brief. Do not self-host Wan as a privacy ritual when the real requirement is "we don't want to think about GPUs."

The Wan 2.7 text-to-video row is the hosted answer. Use it until volume, or a genuine off-box ban, forces the other one.

FAQ

Are the hosted model and the open weights the same quality?

They are the same family. Hosted inference can differ in defaults, safety filters, and how prompt-extend is wired, so a clip will not be a bit-identical twin of a local run. For production purposes they are the same tool. If a look must match a local checkpoint exactly, you are already in self-host territory.

When does self-hosting actually win the money argument?

When Wan is most of the slate, the queue is steady, and the operator cost is already paid. A campaign burst or a mixed-model slate does not get there. Measure a month of hosted Wan credits against a month of GPU plus labour, not against a hardware sticker.

Can I self-host Wan and still use Versely for everything else?

Yes, and that is the usual hybrid: local Wan for the high-volume look, hosted everything else. Do not stand up five local stacks to avoid a credit bill.

Does staying hosted lock me into Versely's Wan forever?

No. The weights are open. Hosted is a convenience and a bill, not a format prison. You can move a prompt and a still to Comfy later. You cannot move a missed deadline back.