According to CoinMeta, Pokee AI has released its new model pokee-isaac 28b, which has a total of 28 billion parameters and supports a context window of up to 10 million token, nearly 10 times that of mainstream models. The current upper limit for mainstream models is around 1 million token. During self-testing, isaac maintained a score of around 95 in long-text tests ranging from 256,000 to 4 million token, and still scored 93.3 after handling a context of 10 million token. The other five comparison models failed to return usable results starting from a context of 2 million token. Pokee attributes its super-long context to its proprietary “non-pure decoder-only architecture”, but the technical report does not disclose the specific architecture or attention mechanism. Pokee AI was founded in 2024 by Zhu Zheqing, former head of applied reinforcement learning at Meta, and has received a $12 million seed round of financing led by Point72 Ventures.