From millions to 100 million tokens, Manta boosts its AI context capacity by 100 times.

Beating AI Express (from Dongcha) – German AI lab Empero has launched a model called Manta, which claims to process up to 100 million tokens at once, 100 times the context window of common million-token models. Empero fed Manta 1,269 annual and quarterly reports of S&P 500 companies, totaling 88 million tokens, then hid 20 secret codes in the documents; the model correctly identified 18 of them. In a second test, 50 highly similar codes were hidden in roughly 10 million tokens, and Manta found 43 correctly. Empero also tested Manta on real AI programming logs: the dataset included 31,450 messages and 7.4 million tokens. The original programming agent compressed its history 13 times due to insufficient context, while Manta processed all data at once and attempted next operations at 5 different positions. Back in 2024, Magic released LTM-2-mini, a model supporting 100 million tokens, claiming its new architecture significantly reduced memory and compute overhead. However, Magic later acknowledged multiple issues when scaling the model’s training, and two years on, the model remains unavailable for independent external testing. Manta has not disclosed its specific architecture, processing speed or usage costs. It remains unclear how the model achieves a 100 million-token context window, nor whether this approach can run stably on larger, more powerful models. Currently, Manta only offers a small number of early trial slots, with testers prohibited from publicly sharing prompts, outputs or test data without authorization.