Researchers from Meta FAIR lab and Meta Superintelligence Labs introduced TextSeal2, a new watermarking method for large language models, on October 29, 2026. The system utilizes a technique called copula mixing to consolidate secret keys and inject noise into the generation process without impacting model speed.
During testing on Llama-3.1-8B, Qwen3.5-9B, and Qwen3.5-27B models, TextSeal2 demonstrated improved detection rates compared to the original TextSeal. For 100-token sequences at a 0.1% false-positive rate, the new method outperformed the predecessor by 8 to 22 percentage points.
The system achieved an 89% detection rate for 200-token outputs on the Qwen3.5-9B model under speculative decoding conditions, while the original TextSeal reached 71% and textGrain achieved 67%. The technology also enables third-party verification through public key sharing.
Meta has released the code and reproducibility scripts for TextSeal2 on GitHub. As of early November 2026, the company has not announced any specific product integrations or public deployments for the technology.
