What does Saient use to run Wan2.1 on Android?

A quantized Wan2.1 T2V 1.3B GGUF transformer, a quantized UMT5-XXL text encoder, and an fp16 VAE, run through a Vulkan-accelerated stable-diffusion.cpp-based engine, with the text encoder and VAE split onto CPU and the diffusion transformer on the phone GPU to fit the memory budget.

What hardware is required?

An Android phone with a Vulkan-capable GPU and roughly 8GB RAM as the target class, plus about 3GB of free storage for the Wan2.1 model pack. It has been developed and verified on a Samsung Galaxy S24 (SM-S921B).

How long does on-device Wan2.1 generation take?

Minutes, not seconds — a 1.3B-parameter video diffusion model is genuinely heavy work for a phone. Start with a short clip (5 frames, 1 step) before trying longer generations.

Run Wan2.1 Locally on Android — On-Device Text-to-Video, No Cloud

Q: Can Wan2.1 actually run locally on a phone?

Yes. Saient runs Wan2.1 T2V 1.3B end-to-end on-device on Android, verified on a stock Samsung Galaxy S24 using the phone's own Vulkan GPU. There is no cloud inference call in the generation path — only the one-time model download uses the network.

Q: Is this the same as a cloud Wan2.1 app with a mobile UI?

No. Many mobile 'Wan' apps are a phone UI in front of a cloud render job. Saient's video engine is a native binary that runs as a real on-device process on the phone, visible as a foreground service doing real local CPU/GPU work, not a remote API call.

Q: Is the engine open source?

Saient's own LLM/SDXL inference engine, Quartz, is open source under MIT at github.com/SaientAI/saient-quartz. The Saient mobile app source, including the Wan2.1 video integration, is at github.com/SaientAI/saient.

The model

What is Wan2.1?

Wan2.1 is an open-weight text-to-video diffusion model. The 1.3B-parameter T2V variant is small enough, quantized, to run on consumer hardware — including, as of Saient mobile, an Android phone GPU, not just a desktop or cloud GPU.

Wan2.1 is developed and released by its own upstream team. Saient does not claim authorship of the model — Saient builds and ships the on-device engine that runs it locally on a phone.

On-device architecture

How Saient runs Wan2.1 on a phone, step by step

Model pack, verified. A quantized Wan2.1 T2V 1.3B transformer (GGUF, Q4_K), a quantized UMT5-XXL text encoder (Q2_K), and an fp16 VAE — about 2.93GB total, each file SHA-256 checked before install.
Chat engine steps aside. Saient's LLM engine (Quartz) unloads first so the phone's memory budget is free for video generation — the two never run at once.
Split backend. The text encoder and VAE run on CPU; only the diffusion transformer — the actual per-step compute — runs on the phone's Vulkan GPU, keeping peak VRAM down.
Disk-paged weights, tiled VAE, hard VRAM caps. Weights are streamed from disk rather than fully memory-resident, and the VAE decodes in tiles, so a multi-gigabyte pipeline fits a phone's shared memory.
Real video, on-device. Output is written as a local .webm file and played back immediately — nothing round-trips to a server.

Full technical write-up, with the exact process arguments and why each one is set the way it is: docs/VIDEO_PIPELINE.md in the Saient mobile repo.

Install guide

How to run Wan2.1 locally on Android with Saient

Clone the app source. git clone https://github.com/SaientAI/saient
Install dependencies. npm ci
Generate the native Android project. npx expo prebuild --platform android --no-install
Build and run on a connected device. npm run android — a native development build is required; Expo Go is not sufficient, since this depends on custom native modules.
Download the Wan2.1 pack in-app. Open the Video tab, download the model pack, and activate it once installed.
Generate. Type a prompt, choose a step/frame preset, and generate — start with the shortest preset (5 frames, 1 step) first.

Saient mobile isn't on an app store yet — building from source is currently the way to run it. See the full README for the production release build path.

Requirements

Supported hardware

✓

Android, Vulkan GPU

Verified — developed and tested on a Samsung Galaxy S24 (SM-S921B), Android 16.

~8GB RAM, ~3GB storage

Target device class. Roughly 3GB free storage for the Wan2.1 pack, plus whatever chat/image models you add separately.

iOS

In progress

Native code needs to be built and exercised on macOS with Xcode before an iOS release.

Source

Where the code lives

github.com/SaientAI/saient — the mobile app, including the Wan2.1 video integration and the on-device pipeline docs.
github.com/SaientAI/saient-quartz — Quartz, Saient's open-source (MIT) LLM chat + SDXL image inference engine.
github.com/Wan-Video/Wan2.1 — the upstream Wan2.1 model.

FAQ

Wan2.1 on Android — questions

Can Wan2.1 actually run locally on a phone?

Yes. Verified end-to-end on a stock Samsung Galaxy S24, using the phone's own Vulkan GPU. No cloud inference call is in the generation path.

Is this the same as a cloud Wan2.1 app with a mobile UI?

No. Several "Wan mobile" apps are a phone UI wrapping a cloud render job. Saient's video engine is a native process running on-device, visible as a foreground service doing real local compute.

Is the engine open source?

Quartz (LLM + SDXL) is MIT-licensed at SaientAI/saient-quartz. The mobile app source, including the Wan2.1 integration, is at SaientAI/saient.

How long does generation take?

Minutes, not seconds — this is real work for a phone. Start with the shortest preset.

Run Wan2.1 on your own phone.

Clone the repo, build it, and generate your first clip on-device.

Get the source Read about the Quartz engine