Meituan open-sources LongCat Avatar 1.5 for multi-speaker talking videos

Most people know Meituan as a food delivery and local services giant, but its AI team has been quietly putting out serious open-source tools. Its latest drop takes single-image talking heads to a level that commercial SaaS tools might find uncomfortable.

Meituan just open-sourced LongCat-Video-Avatar 1.5, arriving only weeks after version 1.0. The premise remains dead simple: feed the model a single reference image and an audio clip, and it spits out a video of the subject speaking naturally.

The full code and weights are already live on GitHub and Hugging Face, including an INT8 quantized version weighing in at 16GB for local deployment.

Natural banter instead of zombie puppets

Most talking-head generators look fine in short cherry-picked clips, but fall apart quickly once the subject starts singing, talking fast, or sharing the screen with someone else. Most models turn multi-person scenes into a bizarre chorus where everyone moves their mouth at once.

Meituan tackled this by swapping out the audio encoder from Wav2Vec2 to Whisper-large, pulling in richer multilingual priors to capture subtle phonemes and rhythm. Face, neck, and torso movements now sync with speech cadences rather than feeling glued on.

The real trick lies in the training recipe. Meituan curated dedicated silent data alongside multi-speaker clips using active speaker detection. The model actually knows who is speaking and who is simply listening, giving the listener natural micro-expressions and gaze shifts instead of fake lip twitches.

To clean up mangled hands and jitters, the team applied frame-level GRPO reinforcement learning. The system handles realistic humans, anime avatars, and stylized characters with consistent styling.

Cutting down to 8 steps for a 15x speedup

Impressive visual quality means little if generating a clip takes all afternoon. Meituan made production throughput the focal point of version 1.5.

Using Distribution Matching Distillation (DMD), the team condensed a typical 50-step diffusion generation pipeline into just 8 steps. Instead of running three heavy models in parallel, the architecture relies on one shared base model with multiple lightweight LoRA adapters.

The result is an estimated 15x bump in inference efficiency, turning out a 10-second video in roughly one minute. In blind evaluations using the EvalTalker benchmark, LongCat 1.5 notched a 65.9% win rate against Kling Avatar 2.0 and 54.3% against HeyGen, keeping frame-dropping issues down to 0.8%.

Why it matters

While major platforms continue locking their best talking-avatar tech behind expensive monthly paywalls, Meituan just handed commercial-grade weights to anyone with a decent GPU.

For independent creators, developers, and small agencies, this dramatically lowers the barrier to video automation. Producing product explainers, localization dubs, or even multi-character conversational shorts no longer requires pricey enterprise contracts or dealing with glitchy, wooden open-source alternatives. A 16GB INT8 build makes local prototyping practical.

This release also raises the baseline for the entire avatar industry. When single-portrait synthesis, multi-speaker logic, and rapid 8-step inference are available openly, proprietary vendors will have to offer much deeper workflow value beyond simple lip-syncing.

That said, real-time interactive avatars are still out of reach here. A minute of compute for ten seconds of video is fast for batch processing, but not fast enough for live streaming. We think the immediate payoff lies in high-volume, automated marketing assets and scalable e-commerce production lines.