Meta drops five new AI models spanning multimodal, music, and watermarking

Meta's FAIR team just released five practical AI models and research tools to the public, spanning mixed-modal generation, precision-controlled music synthesis, multi-token language modeling, and synthetic voice watermarking.

Instead of chasing raw parameter size, this drop tackles specific engineering bottlenecks: native multimodal input-output, deterministic control in audio creation, inference efficiency, and real-time deepfake audio tracking.

What these tools can do

Chameleon is a mixed-modal model family. While typical multimodal models only handle one-way tasks like text-to-image or image captioning, Chameleon takes any arbitrary mix of text and images as input and generates any mix of text and images as output. You can feed it photos alongside narrative notes and receive an interspersed multimodal document.

JASCO takes text-to-music generation beyond simple prompt guesswork. Previous models like MusicGen rely almost purely on text descriptions, making precise rhythm impossible. JASCO lets you feed in chords or beat patterns alongside text prompts, merging symbolic controls and audio inputs so creators can keep musical structure tightly governed without sacrificing audio quality.

Under the hood: speedups and 485x faster watermarking

The drop also includes foundational training tweaks and trust-and-safety tooling.

The multi-token prediction release tackles the traditional inefficiency of predicting one next word at a time. By training language models to predict multiple future tokens simultaneously, it significantly speeds up code completion workflows. Meta has shared the pretrained weights specifically tuned for code completion tasks under a research license.

AudioSeal stands out as the first localized watermarking system built specifically for AI-generated speech. Unlike older techniques that require heavy decoding algorithms over an entire track, AudioSeal can pinpoint the exact synthetic slices inside a long audio clip. It runs up to 485 times faster than existing methods, making it practical for real-time and massive-scale moderation pipelines.

Meta also released evaluation code and a dataset of over 65,000 human annotations to help developers audit geographical and cultural representation in text-to-image systems.

Licenses and availability

The licensing terms depend on the tool. AudioSeal is released under a commercial license, ready for immediate deployment in commercial platforms and safety pipelines.

Chameleon's key components and the multi-token code completion weights are distributed under non-commercial, research-only licenses. The geographic disparity evaluation codebase and survey dataset are fully open to the community for benchmark auditing.

Why it matters

If you build audio tools or moderate user-generated content, AudioSeal is an instant win: it offers production-grade synthetic speech detection without the massive latency and compute overhead of older watermarking tools.

We see this as another smart tactical move by Meta. By handing out tricky architectural pieces like mixed-token prediction, native multimodal fusion, and watermarking rails, Meta avoids empty benchmark chest-thumping and instead lets builders do the legwork of stress-testing and standardizing its architectures across the open ecosystem.