Pokee AI Releases Pokee-Isaac 28B with 10M-Token Context Window

The artificial intelligence landscape welcomes Pokee-Isaac 28B, an advanced agentic model developed by Pokee AI designed to handle an enormous context window of 10 million tokens. The model is offered as a licensed solution through an OpenAI-compatible developer API, with deployment options tailored for virtual private clouds (VPCs), on-premises environments, or local hardware devices.

Regarding infrastructure support, the launch announcement details Day-0 compatibility with vLLM and SGLang, alongside single-GPU serving requirements beginning with an RTX 4090 or equivalent hardware. However, the research team published measurements derived exclusively from a B200-class GPU, meaning consumer-GPU deployment claims should be viewed as vendor guidance rather than benchmarked outcomes.

Stock photo for illustration only, not from the actual event

When evaluated across rigorous benchmark suites, the Pokee-Isaac model delivers exceptional performance metrics. According to the RULER evaluation framework, the model maintains an accuracy score remaining above 93.3% across every tested length, concluding at 93.3% at the 10-million token mark. By comparison, baseline models such as GPT-5.6 Luna and Gemini 3.5 Flash Lite track closely up to 512K before encountering context-overflow errors at 1M.

Furthermore, in the MRCR v2 evaluation utilizing 8 needles, Isaac achieves scores of 0.607, 0.743, and 0.500 at context lengths of 256K, 512K, and 1M respectively, expanding its performance margin over Gemini from 0.133 to 0.295 throughout the evaluation sweep. In DTAP red-teaming assessments, Isaac records the lowest direct attack success rate at 36.0, indirect at 35.2, and combined at 35.6, accompanied by an 82.5% benign success rate, noting that baseline models executed under the stock runner while Isaac utilized the Pokee harness.