From a million to 100 million tokens, Manta increases AI context capacity by 100 times.

Beating AI News Flash: German AI lab Empero has released a model called Manta, which claims to read 100 million tokens at once—100 times that of the commonly seen million-token context models.

Empero crammed a total of 1,269 annual and quarterly reports from S&P 500 companies into Manta in one go—a full 88 million tokens. Then they hid 20 secret codes in these documents and had the model find them one by one. It found 18 correctly. The team tried again, hiding 50 very similar secret codes in about 10 million tokens, and Manta found 43 correctly.

Empero also tested it with real AI coding records. This record had 31,450 messages and 7.4 million tokens; the original coding agent, due to insufficient context, compressed its history 13 times. Manta read it all in one pass and attempted to continue writing the next step at 5 different positions.

Back in 2024, Magic had released LTM-2-mini, which supported 100 million tokens, and claimed that a new architecture greatly reduced memory and compute overhead. But Magic later admitted that the model encountered many problems when scaling up training. Two years on, this model is still not publicly available for independent testing.

Manta has also not disclosed its specific architecture, processing speed, or usage cost, and it remains unclear how it achieves a 100-million-token context, let alone whether this approach can run stably on larger, more powerful models.

Currently, Manta only offers a small number of early trial slots, and trial users may not publicly share prompts, outputs, or test data without permission.