Alibaba's Wan 2.7 text-to-video model reaches Replicate with 1080p output and synced audio

What Wan 2.7 does

According to a beginner's guide published on dev.to by AImodels.fyi, wan-2.7-t2v is a text-to-video generation model maintained by the Wan-Video team and built on Alibaba's Wan 2.7 architecture. It produces clips of two to 15 seconds — five seconds by default — at resolutions up to 1080p, with 720p and 480p options also available. The default aspect ratio is 16:9, with other ratios selectable through a parameter.

Under the hood, the guide describes a diffusion transformer architecture trained on billions of images and videos, paired with a custom variational autoencoder that can encode and decode full-HD footage while keeping temporal information intact. The distinguishing feature is audio: wan-2.7-t2v can synthesize a matching soundtrack automatically, or synchronize against a file the user uploads in WAV or MP3 format (three to 30 seconds, up to 15 MB). The guide calls it one of the few open-source models that handles picture and sound coherently in a single pass.

Practical API access

The model is hosted on Replicate, where the most recent version was deployed on April 2, 2026. Weights and inference code are also distributed through Hugging Face and ModelScope under the repository's license, so developers can self-host rather than depend on the API.

The input surface is compact. A text prompt is the only required field, joined by an optional negative prompt for steering away from unwanted content, plus controls for duration, resolution and aspect ratio. A seed between 0 and 2147483647 allows reproducible runs, which the guide highlights as useful for generating A/B variations of a clip from slightly different prompts. Prompt expansion is enabled by default so short prompts get fleshed out automatically, though the guide notes this increases latency.

Where the guide says it fits

AImodels.fyi points at several use cases: turning marketing copy into promotional footage, building educational or tutorial videos that combine visuals with narration, producing cinematic B-roll for streamers and podcasters, and generating multiple variations for testing. It can also render text inside videos in both Chinese and English, though the guide flags the quality of that text rendering as unpredictable.

Against earlier Wan releases, the comparison is straightforward. Version 2.7 improves on 2.5 with better audio-video synchronization and slightly stronger visuals, while 2.5 remains the pick when inference speed matters most. It also supersedes 2.6 with cumulative gains in prompt understanding and output stability, though teams with configurations tuned for 2.6 may have reason to stay put. The smaller wan-2.1-1.3b model is far cheaper to run — the guide cites 8.19 GB of VRAM, making it viable on consumer GPUs — while the fast variant of 2.5 targets sub-minute generation times. Wan 2.7 also differs from wan-2.5-i2v, which animates a supplied starting image rather than generating from scratch.

Constraints to plan around

The 15-second duration ceiling rules out long-form work. The guide reports that quality can drop at very high resolutions, noting that the 1.3B variant is more stable at 480p, so the 1080p setting on the Replicate deployment may produce inconsistent results. Camera movements specified in prompts can come out jittery or unrealistic, audio sync works best within the three-to-30-second window, and inference is compute-heavy enough that a single five-second clip at full quality takes considerable time. Negative prompts help but cannot guarantee that unwanted elements stay out, and the training data composition is not fully documented, which leaves open the possibility of unknown biases in outputs.

Why it matters

A capable open-weights text-to-video model with audio in one pass, reachable through a simple hosted API, lowers the barrier for developers who want to prototype generative video without standing up their own GPU infrastructure. The practical trade-offs — short clips, slow inference, variable 1080p quality — are real, but the combination of a permissive licensing posture, multiple distribution channels and a compact parameter set makes Wan 2.7 a serious option alongside faster, lower-quality variants for teams building video generation into products.