Ranked #1 Local Video Generation — Your GPU, Your Director's Chair
Alibaba Cloud (Tongyi Lab)

Wan 2.7

The open-weight video model that learned to think before it shoots. Wan 2.7 is Alibaba's biggest leap yet — a 27B Mixture-of-Experts model that plans your scene, syncs audio natively, and gives you First and Last Frame control, all under the most permissive license in AI. Apache 2.0. No asterisks.

Updated April 2026 Open WeightApache 2.0Text-to-Video
8.9out of 10
Official Website
Best for

The open-weight video model that learned to think before it shoots. Wan 2.7 is Alibaba's biggest leap yet — a 27B Mixture-of-Experts model that plans your scene, syncs audio natively, and gives you First and Last Frame control, all under the most permissive license in AI. Apache 2.0. No asterisks.

Why It Wins

Thinking Mode plans scene structure before generation, dramatically reducing motion drift and generic compositions. Native audio sync in a single pass. First and Last Frame control for precise transitions. Up to 9 multimodal reference inputs for character consistency. Instruction-based video editing. 27B MoE architecture. Fully available on Alibaba Cloud Model Studio, fal.ai, and WaveSpeedAI. Apache 2.0.

Watch out

The Thinking Mode that makes it special also makes it slower — planning before generating adds time. No sign-up-and-go cloud API; still requires technical comfort. The 27B scale means serious hardware requirements for local deployment. Documentation remains primarily Chinese-first.

01

What It Actually Is

Wan 2.7 is what happens when Alibaba asks: what if the AI actually thought about your video before making it?

Every previous version of the Wan series — and most AI video generators generally — work the same way: receive prompt, generate video, hope for the best. Wan 2.7 breaks that pattern. Its Thinking Mode interprets your prompt, plans the scene structure and narrative arc, considers composition and motion dynamics, and then begins rendering. The output is less generic, less drifty, and more intentional. It’s the difference between handing a script to a camera operator and handing it to a director.

The technical foundation is a 27B Mixture-of-Experts architecture with a Diffusion Transformer and Full Attention mechanism — designed specifically to process spatial and temporal relationships simultaneously. This isn’t just a scaled-up version of earlier Wan models; it’s a different approach to how video generation should work.

The practical upgrades are equally significant. First and Last Frame control lets you define the exact start and end of a clip — critical for maintaining continuity across scenes. Up to 9 multimodal reference inputs (images, clips, audio) keep characters, props, and styles consistent without manual stitching. Native audio sync generates sound in the same pass as the visuals. And instruction-based editing means you can refine outputs by describing the change you want, rather than regenerating from scratch.

The Apache 2.0 license remains the Wan series’ defining commitment. Not “open with enterprise restrictions.” Not “free for personal use.” Apache 2.0 — the same license that governs the infrastructure half the internet runs on. Use it commercially, modify the weights, build products, sell the output. Zero asterisks.

The community that made Wan 2.1 the ComfyUI standard is already moving to 2.7. The ecosystem will follow — it always does.

02

Strengths and honest limitations

Key Strengths

  • Thinking Mode — plans before it generates: Wan 2.7’s defining innovation. Before rendering a single frame, the model interprets your prompt, plans the scene composition, and structures the narrative arc. The result is dramatically fewer generic compositions, less motion drift across frames, and higher intent fidelity. It’s the difference between a camera operator and a director.
  • First and Last Frame control: Define the exact starting and ending frames of your video. This gives you precise control over transitions, character positions, and scene continuity — something almost no other open-weight model supports at this quality level.
  • Native audio sync: Audio is generated alongside video in the same pass — not bolted on after the fact. Dialogue timing, ambient sound, and music emerge as a unified composition with the visuals.
  • Reference consistency up to 9 inputs: Provide up to 9 multimodal reference images or video clips to anchor character identity, style, and props across scenes. Multi-scene consistency — the hard problem of AI video — is now achievable in an open-weight model.
  • Instruction-based video editing: Describe the change you want in natural language; Wan 2.7 modifies the existing video accordingly. Edit without re-generating from scratch.
  • Apache 2.0 — genuinely no restrictions: Commercial use, weight modification, product building, output selling — all permitted with no revenue caps and no attribution beyond the license file. The most commercially friendly open-weight video model at this capability level.

Honest Limitations

  • Thinking Mode adds generation time: The planning pass that makes outputs better also makes them slower. If speed is the priority, LTX Video remains the faster choice at comparable quality.
  • No official sign-up-and-go API: You run it locally or via community-hosted endpoints (fal.ai, WaveSpeedAI). For non-technical users, the setup barrier is real.
  • 27B scale demands serious hardware: Local deployment at full quality requires GPU memory well beyond consumer cards. Cloud GPU rental or platform API access is the practical path for most users.
  • Chinese-dominant documentation: Official docs and many community resources are primarily in Chinese. English community guides exist but lag updates.
  • Newer — smaller community than Wan 2.1: The Wan 2.1 ComfyUI ecosystem is vast. Wan 2.7 workflows and custom nodes are growing but not yet at the same scale.
03

Benchmark Snapshot

Architecture — 27B MoE (40B active params)

Mixture-of-Experts design with ~753B total parameters, ~40B active per token. DiT with Full Attention for spatial and temporal processing. IndexShare sparse attention reduces compute at extreme context lengths.

Resolution — Up to 4K

Supports up to 1080p and 4K output depending on the platform and compute tier, with clip durations from 5 to 15 seconds.

Reference capacity — Up to 9 multimodal inputs

Accepts a combination of images, video clips, audio, and style references to maintain character, prop, and aesthetic consistency across multi-scene generations.

04

The Verdict

Wan 2.7 is the moment open-weight video generation grew up. The Thinking Mode isn’t a gimmick — it’s the difference between an AI that responds to your prompt and one that interprets it. First and Last Frame control, native audio, up to 9 reference inputs, instruction-based editing: these are capabilities that were either absent or unreliable in open models until now. And all of it ships under Apache 2.0, meaning you own what you build with it, completely. The trade-off is complexity and compute — this is not a point-and-click tool. But for creators who want cinematic control without subscription fees or usage restrictions, Wan 2.7 is the most capable open-weight video model in existence.

05

Frequently Asked Questions

Keep comparing
See full ranking
#2

LTX Video 2.3

Lightricks

The speed demon of local video generation — and the only local model that generates synchronized audio and video in a single pass. Lightricks built a 22-billion parameter model that produces 1080p video with dialogue, music, and sound effects baked in, not bolted on. Licensed training data from Getty and Shutterstock means less copyright anxiety.

8.5 Editorial score
Read review