Deploy Wan 2.7 on GPU Cloud: AI Video Generation Setup (2026)

Wan 2.7 is the version everyone's searching for right now, and it's also the version you can't run yourself. Alibaba's Tongyi Lab shipped it API-only in April 2026, the same pattern it set with Wan 2.5 and Wan 2.6. If you came here to "deploy Wan 2.7," the honest answer is that you deploy Wan 2.2, the last version with public weights, and this guide covers exactly how: VRAM requirements, ComfyUI and diffusers setup on Spheron H100 or B200, and what a clip actually costs you compared to the Wan 2.7 API. For the predecessor deployment guide this one builds on, see deploying Wan 2.5 on GPU cloud and the original Wan 2.1/2.2 ComfyUI setup.

What's New in Wan 2.7: Thinking Mode, Frame Control, 9-Grid Input, and Long Prompts

Wan 2.7 (Wan2.7-Video) is Tongyi Lab's latest video generation suite. The text-to-video component went live April 3, 2026, with the full I2V, R2V, and editing suite following April 8-9. A Tongyi Lab spokesperson framed the release around control: "Wan 2.7 marks a major leap forward in controllable generative AI. We are delivering tools that combine creative freedom with precise control like never before."

The feature list backs that up. Four capabilities separate Wan 2.7 from everything that came before it in the Wan series:

• Thinking Mode. Instead of decoding a prompt directly into pixels, the model first builds a compositional plan, how elements relate in space and time, what the narrative logic of the sequence should be, and only then starts generating frames. The tradeoff is speed: a 5-second clip takes roughly 4 minutes to generate via the API, and a 15-second 1080p generation can run past 10 minutes.
• First/last-frame control (FLF2V). You supply a starting image and an ending image, and the model fills in the motion between them. This is built for product shots, 360-degree reveals, and storyboard tests where you already know your bookends.
• 9-grid input. Feed the model a 3x3 grid of product photos and it turns the set into a coherent video sequence, aimed squarely at e-commerce catalogs that need motion content at scale.
• Long prompts. The image-led API path reportedly accepts prompts up to 5,000 characters with negative prompts up to 500 characters, well beyond what earlier Wan versions handled. Treat that figure as a secondary-source estimate rather than an Alibaba-published spec; it hasn't been confirmed on a primary Alibaba page.

Underneath those four, Wan 2.7 is really four models bundled into one suite: text-to-video (up to 1080p, with lip-sync), image-to-video, reference-to-video (for multi-character consistency across a clip), and a natural-language video editing model that takes an existing clip and a text instruction. It also generates native audio alongside video, with the option to preserve or regenerate the track, and supports 2-15 second clips (10 seconds max in R2V mode) at 720p or 1080p across five aspect ratios: 16:9, 9:16, 1:1, 4:3, and 3:4.

That's a genuinely broader feature set than Wan 2.2. It's also entirely unavailable to you as a self-hosted deployment.

Wan 2.7 Is API-Only: Why "Deploy Wan 2.7" Actually Means Deploying Wan 2.2

Every platform offering Wan 2.7 offers it exactly one way: as a hosted API call. fal.ai lists it as a cloud endpoint with no open-weight or self-hosting option mentioned anywhere on the model page. Together AI serves it through the same serverless API pattern it uses for every other model on its platform, starting with the text-to-video endpoint. Neither company gives you a checkpoint to download.

The GitHub evidence backs that up directly: there's no Wan-Video/Wan2.7 repository. The URL returns a 404. Compare that to Wan 2.1 and Wan 2.2, which both have public repos with published weights on HuggingFace. Tellers.ai, reviewing the release in April 2026, put it plainly: "the weights have not been released" and "access is API-only," calling it a strategic reversal from the fully open Wan 2.1.

That reversal isn't unique to 2.7. It's the pattern for every Wan release since 2.2.

What Alibaba Has (and Hasn't) Open-Sourced Across the Wan Series (2.1 through 2.7)

Wan 2.2 is the last publicly-weighted release, and it's been the last one for a full year of Wan version bumps. If you need self-hosted video generation on your own GPU, under your own cost structure, with no per-second billing, Wan 2.2 is what you deploy. Everything from 2.5 onward is a hosted product, not infrastructure you control.

For the fuller architecture comparison between Wan 2.1 and Wan 2.2, including the Mixture-of-Experts change and training data scale, see the Wan 2.1/2.2 deployment guide. For where Wan 2.5 sits in this same pattern, see deploying Wan 2.5 on GPU cloud.

GPU and VRAM Requirements for Video Generation Today (Wan 2.2, the Latest Open-Weight Version)

Wan 2.2 is a Mixture-of-Experts diffusion transformer: 27B total parameters, 14B active per denoising step, trained on 65.6% more images and 83.2% more videos than Wan 2.1. VRAM requirements carried over unchanged from 2.1, so the hardware math below applies to either version.

Going from 480p to 720p increases spatial token count by 2.25x, but attention memory scales quadratically with token count, so VRAM grows roughly 2-3x for that same resolution jump. That's the number that determines your GPU tier, not the pixel count on its own.

GPU selection:

• H100 PCIe on Spheron (80GB, $2.01/hr on-demand, $1.67/hr spot): the minimum viable GPU for Wan 2.2 14B at 720p with FP8 quantization. Margin is tight; confirm headroom with nvidia-smi on your first job before scaling to production.
• H100 SXM5 (80GB, $2.54/hr on-demand): same VRAM ceiling as PCIe, but 3.35 TB/s memory bandwidth versus 2 TB/s cuts generation time roughly 25% on attention-heavy workloads.
• H200 SXM5 (141GB, $5.92/hr on-demand, $3.31/hr spot): the right call for 10-second 720p clips or any workload where you don't want to think about VRAM margin.
• B200 GPU rental on Spheron (192GB, $7.50/hr on-demand, $5.34/hr spot): eliminates VRAM concern entirely, including FP16 at 720p and 10-second clips. Faster memory bandwidth also cuts wall-clock generation time versus H100.

Multi-GPU note: a single video generation job doesn't benefit from tensor parallelism in standard ComfyUI or diffusers setups. One job runs on one GPU. You scale throughput by running more GPU instances in parallel, each on an independent job, not by splitting one clip across GPUs.

Why Wan 2.7's 1080p Ceiling Doesn't Exist On-Prem Yet

Wan 2.7 generates natively at 1080p. Wan 2.2, the version you can actually run yourself, doesn't officially support 1080p at all, its published resolution ceiling is 720p. That's not a licensing restriction, it's a training and architecture gap: Alibaba hasn't released a self-hostable Wan checkpoint trained for 1080p output.

Practically, this means the API-vs-self-hosted comparison later in this post isn't apples to apples at the top resolution tier. If your product requires native 1080p output today, the API is your only route regardless of cost. If 720p is acceptable, either against a downstream upscaler or as your ship resolution, Wan 2.2 self-hosted is on the table.

Deployment Walkthrough: Wan 2.2 with ComfyUI and Diffusers on Spheron H100/B200

This is the same infrastructure covered in the Wan 2.5 deployment guide, repeated here for anyone landing on this post directly. If you already have a Wan 2.2 pipeline running from that guide, skip to the cost section below.

Step 1: Launch an instance

Provision an H100 PCIe instance on Spheron for 720p work at standard clip lengths, or a B200 instance if you need 10-second clips or want VRAM margin (see GPU selection above). Choose Ubuntu 22.04. Don't expose port 8188 publicly, ComfyUI ships with no built-in authentication; you'll reach it through an SSH tunnel.

Step 2: Deploy ComfyUI via Docker

SSH into the instance and run:

IMAGE=ghcr.io/ai-dock/comfyui:latest-cuda
docker pull $IMAGE

docker run -d \
  --name comfyui \
  --gpus all \
  --ipc=host \
  -p 127.0.0.1:8188:8188 \
  -v ~/comfyui-models:/opt/ComfyUI/models \
  -v ~/comfyui-output:/opt/ComfyUI/output \
  $IMAGE

The -v mounts persist model weights and output across container restarts. --ipc=host is required for PyTorch's shared memory. Binding to 127.0.0.1 keeps ComfyUI off the public interface.

Step 3: Install ComfyUI-WanVideoWrapper

docker exec -it comfyui bash
cd /opt/ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt
exit
docker restart comfyui

Check the WanVideoWrapper GitHub if you hit dependency errors; the requirements file changes often between releases.

Step 4: Download Wan 2.2 weights

pip install huggingface_hub

huggingface-cli download Wan-AI/Wan2.2-T2V-A14B \
  --local-dir ~/comfyui-models/wan-t2v-14b

# For image-to-video:
# huggingface-cli download Wan-AI/Wan2.2-I2V-A14B \
#   --local-dir ~/comfyui-models/wan-i2v-14b

The 14B T2V checkpoint is roughly 69GB total (DiT weights plus the T5 text encoder and VAE). Budget 30-90 minutes to download depending on connection speed.

Step 5: Access via SSH tunnel and generate

ssh -L 8188:localhost:8188 user@your-server-ip

Open http://localhost:8188, load a Wan 2.2 workflow JSON (comfyworkflows.com or the WanVideoWrapper GitHub have working examples), set your prompt, and queue the job. On H100 PCIe, expect roughly 4 minutes for a 480p 5-second clip and 10-12 minutes at 720p.

For teams building a programmatic pipeline instead of a UI, the diffusers library gives you a Python-native path:

import torch
from diffusers import WanPipeline

pipe = WanPipeline.from_pretrained(
    "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

video = pipe(
    prompt="A slow-motion close-up of water droplets falling onto a dark stone surface",
    num_frames=25,
    height=480,
    width=832,
    num_inference_steps=40,
    guidance_scale=5.0,
).frames[0]

import imageio
imageio.mimsave("output.mp4", video, fps=24)

Keep the pipeline resident in GPU memory between requests if you're wrapping this in an API. The 14B model takes 30-60 seconds to load; reloading per request halves your effective throughput. Full setup detail, including the ComfyUI on GPU cloud node-based workflow and FastAPI wrapping pattern, is in the Wan 2.5 deployment guide.

Pricing fluctuates based on GPU availability. The prices above are based on 26 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Cost per Clip: Wan 2.7 API Pricing vs Self-Hosted Wan 2.2 on Spheron

fal.ai prices every Wan 2.7 endpoint at a flat $0.10 per second of generated video, regardless of resolution. A 10-second 1080p clip costs $1.00. A 5-second clip costs $0.50. Apimart.ai's breakdown of the same model shows a resolution-tiered structure instead, roughly $0.0664/second at 720p and $0.1096/second at 1080p, so shop the specific endpoint you're using; flat-rate and tiered pricing aren't interchangeable.

Here's what the same clip lengths cost self-hosted on Wan 2.2, run on Spheron:

Pricing fluctuates based on GPU availability. The prices above are based on 26 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

The 5-second case is straightforward: self-hosted Wan 2.2 on H100 PCIe on-demand beats the Wan 2.7 API outright, and spot pricing shaves off another ~17% (H100 PCIe spot runs $1.67/hr against $2.01/hr on-demand) if your workload tolerates preemption. The 10-second case is where it gets honest. Pushing Wan 2.2 to 10 seconds at 720p needs H200-class VRAM headroom, and at that GPU's hourly rate, the self-hosted clip can cost more than the $1.00 API price for the same duration. Neither path wins in every scenario, and resolution matters too: the API gets you to native 1080p, Wan 2.2 doesn't.

The math that actually decides this for you is volume and idle time, not per-clip cost alone. A rented GPU bills whether or not it's generating a clip. At low, bursty volume, the API's zero idle cost wins even where its per-clip rate looks worse on paper. Past roughly 40-60 clips a day running consistently, a dedicated instance's lower marginal cost per clip starts to outweigh the hours it sits idle between jobs, and self-hosting takes over as the cheaper path. For the fuller breakeven math and production queuing pattern (job queue, GPU worker pool, object storage for output), see the Wan 2.5 deployment guide's cost section, and for how Wan pricing stacks up against nine other inference APIs, see Fal.ai alternatives for video generation.

For the broader video model landscape beyond Wan, including HunyuanVideo and LTX-Video comparisons, see the AI video generation GPU guide. For VRAM sizing across the full 2026 video AI stack, see GPU cloud for video AI 2026. If your use case is animating a fixed reference image rather than generating from scratch, the image-to-video deployment guide covers Wan 2.2 I2V alongside LTX-Video and Hunyuan Video Avatar. And if you want a Wan 2.2 checkpoint tuned to your own product or brand rather than the base model, see the Wan LoRA training cost guide for musubi-tuner setup and real per-run dollar figures.

FAQ

Can I self-host Wan 2.7?

No. It shipped API-only in April 2026, with no GitHub repo, no HuggingFace weights, and no self-hosting path through any platform currently offering it. Wan 2.2 is the version you deploy if you need to run video generation on your own hardware.

What's the difference between Wan 2.7 and Wan 2.2?

Wan 2.7 adds Thinking Mode, first/last-frame control, 9-grid product input, native 1080p, native audio, and long prompts. Wan 2.2 caps at 720p, video only, no Thinking Mode or frame control. The tradeoff is access: Wan 2.2 weights are public, Wan 2.7's aren't.

How much VRAM does Wan 2.2 need?

Roughly 40-48GB at 480p with FP8, and 65-80GB at 720p. H100 PCIe (80GB) is the practical minimum for 720p; H200 or B200 gives real headroom, especially for 10-second clips.

Is Wan 2.7 or self-hosted Wan 2.2 cheaper?

Depends on clip length. A 5-second 720p Wan 2.2 clip on H100 PCIe on-demand runs about $0.34-0.40, cheaper than the $0.50 you'd pay through fal.ai's Wan 2.7 API at the same length. A 10-second 720p clip needs H200-class hardware for self-hosting, and at that point the self-hosted cost can run higher than the API's $1.00. Above roughly 40-60 clips a day, self-hosting usually wins on total cost; below that, the API's zero idle cost usually wins.

Does Wan 2.7 support audio generation?

Yes, with the option to preserve the original audio or regenerate it. Wan 2.2 doesn't generate audio.

Wan 2.2 production workloads run today on Spheron's H100 and B200 bare-metal instances, from $2.01/hr on-demand with no per-second API fees and full root access.

Spheron H100 for Wan 2.2 → | B200 GPU rental → | View all GPU pricing →