Pokee AI has unveiled a significant advancement in the field of large language models with the launch of Pokee-Isaac 28B, a 28-billion-parameter text-only foundation model designed to operate within customer environments. This new model stands out for its impressive 10-million token context window, a capability that positions it among the most advanced models in terms of handling long-form inputs.
Performance and Technical Specifications
The model demonstrates exceptional performance on key benchmarks, scoring 93.3% on the RULER test at 10 million tokens. Notably, all baseline models in its comparison group achieve a score of 0.0 beyond the 2-million token threshold, highlighting Pokee-Isaac's superior long-context capabilities. Additionally, it leads the BFCL v4 benchmark with a score of 70.94 and ranks second on Terminal-Bench 2.1. The model’s efficiency is also impressive, achieving a prefill rate of 137,200 tokens per second at full context on a single B200 chip, with a decoding speed of approximately 335 tokens per second.
Deployment and Accessibility
Unlike many models where weights are publicly available, Pokee-Isaac 28B’s weights are not published. Instead, deployment is restricted to secure environments such as Virtual Private Clouds (VPC), on-premises installations, or on-device setups, making it ideal for organizations with strict data privacy requirements. The pricing model is structured at $0.15 or $1.00 per million tokens, depending on the deployment method. This approach underscores Pokee AI’s focus on enterprise-grade security and control, catering to use cases where data sovereignty is paramount.
The release of Pokee-Isaac 28B marks a pivotal step in the evolution of agentic models that can operate with extended context while maintaining performance and privacy. As organizations seek more secure, scalable, and capable AI solutions, this model may serve as a compelling option for enterprise deployments.
