Empero released Manta and stated that it supports a 100000000 Token context.

Article

From millions to 100000000 Token, Manta increases AI context capacity by two orders of magnitude

Dongcha Beating AI Newsflash: German AI lab Empero released a model called Manta, claiming it can read in 100000000 Token at once, which is 100 times the current common million Token context models.

Empero stuffed all 1269 annual and quarterly reports from S&P 500 companies into Manta at once, totaling 88000000 Token. Then they hid 20 passcodes within these files and asked the model to find them one by one. The result was that 18 were found correctly. The team tried again, hiding 50 very similar passcodes in approximately 10000000 Token, and Manta found 43 correctly.

Empero also tested with real AI programming records. This record contained 31,450 messages and 7400000 Token; the original programming agent compressed historical records 13 times due to insufficient context. Manta read it all at once and attempted to continue writing the next step operations at 5 different positions.

As early as 2024, Magic had released LTM-2-mini supporting 100000000 Token, claiming to significantly reduce VRAM and computational overhead through a new architecture. However, Magic later admitted encountering many issues when scaling up the model's training. Two years have passed, and this model has still not been made public for independent external testing.

Manta currently does not disclose specific architecture, processing speed, or usage costs. It remains unclear how it achieves 100000000 Token context, nor whether this method can run stably on larger, stronger models.

Currently, Manta only opens a small number of early trial slots. Trial users are not permitted to publicly share prompts, outputs, or test data without permission.

Original link https://m.theblockbeats.info/flash/376278

Key points

German AI lab Empero released a model named Manta.
Empero claims that Manta can read in 100000000 Token at a time.
In a test containing 1269 S&P 500 company annual and quarterly reports and about 88000000 Tokens, Manta correctly found 18 of the 20 secret codes.
In a test of hiding 50 similar codes among approximately 10000000 Tokens, Manta correctly identified 43.
An AI programming record containing 31450 messages and 7400000 Tokens was read in one go by Manta.
Manta currently offers only a limited number of early access slots; prompts, outputs, and test data may not be shared publicly without permission.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Context capacity can be understood as the amount of text a model can process at one time. Tests reported indicate that Manta can find information or continue operations within ultra-long documents and programming records, but these results primarily come from tests conducted by the publisher.

Why it matters to readers

If the results are reproducible, larger context capacity may help process long documents and lengthy AI programming records. It is currently impossible to judge its actual speed, cost, or stability based on this.

This indicates that the capabilities of AI models depend not only on answer quality but also on how much content they can read at once; however, promotional data requires verification through independent testing.

Risks and unknowns

The specific architecture of Manta has not been disclosed.
Processing speed and usage costs have not been disclosed.
It is unclear whether the relevant test results can be independently reproduced by external teams.
It is unclear whether this method can run stably on larger, more powerful models.
Trial users may not share prompts, outputs, and test data publicly without permission.

Empero released Manta and stated that it supports a 100000000 Token context.

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center