ChainThink reports that on August 5, AI Agent startup Pokee AI launched Pokee-Isaac 28B, a model with 28 billion parameters and a context window of 10 million tokens—nearly ten times the typical 1 million token limit of mainstream models.
According to its own testing, Isaac maintained a score of around 95 throughout the RULER long-context test range from 256,000 to 4,000,000 tokens, and still scored 93.3 when pushed to 10,000,000 tokens.
The other five benchmark models did not return usable results starting from 2 million tokens.
Pokee attributes this capability to its proprietary "non-pure decoder-only architecture," but the technical report does not disclose the specific architecture, attention mechanism, or memory management scheme, leaving the implementation method unclear to outsiders.
Pokee AI was founded in 2024 by Zhu Zheqing, former head of applied reinforcement learning at Meta. The company raised a $12 million seed round led by Point72 Ventures, with participation from Qualcomm Ventures, Samsung NEXT, and others.
