MultiTalk
Meituan's open-source audio-driven video model that animates one or several people in conversation, up to 15 seconds, at 480p or 720p.
About
MultiTalk ('Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation', a NeurIPS 2025 paper from Meituan's MeiGen-AI team) is built for a gap most talking-head models ignore: conversations. It drives single- or multi-person video from audio streams, supports text prompts to direct the virtual humans, handles cartoon characters and singing, and renders at 480p or 720p in arbitrary aspect ratios with clips up to 15 seconds. The project has stayed actively maintained, adding INT8 quantization, SageAttention, faster CFG strategies, and FusionX/lightx2v LoRA acceleration that cut inference to a handful of steps.
The same team has since released InfiniteTalk for infinite-length video dubbing and the LongCat-Video-Avatar series, so MultiTalk now sits inside a growing family of open avatar models. Everything is Apache-2.0 with code, weights, and Gradio demos public, and the repo (3k stars) was still being updated in October 2026. There are no published benchmark scores, so its standing rests on community adoption and the NeurIPS peer review rather than numbers.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published benchmarks; NeurIPS 2025 paper with author-run comparisons
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: NeurIPS 2025 paper from Meituan MeiGen-AI; GitHub MeiGen-AI/MultiTalk (3k stars, Apache-2.0 confirmed via GitHub API, updated Oct 2026); sibling releases InfiniteTalk and LongCat-Video-Avatar show sustained investment. Strengths: rare multi-person conversational focus, permissive license, active maintenance, peer-reviewed. Docked for: (1) no published benchmark scores; (2) official repo URL cross-referenced from fork mirrors rather than a first-party announcement page. License: Apache-2.0, confirmed via GitHub API.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Meituan MeiGen-AI; NeurIPS 2025 paper 'Let Them Talk'; released Jun 2025
- · Audio-driven single- and multi-person conversational video; 480p/720p up to 15s
- · 3k stars; INT8 quantization, SageAttention, FusionX/lightx2v acceleration added
- · Sibling releases: InfiniteTalk, LongCat-Video-Avatar; no published benchmarks