Pokee AI Releases Pokee-Isaac 28B: 10M-Token Context Model

Pokee AI released Pokee-Isaac 28B, a text-only foundation model with 28 billion parameters and a 10 million token context window, engineered to operate within a customer's data boundary. This capability is crucial for regulated industries, public sector institutions, and on-device applications where data cannot leave the premises. Traditional cloud-based large language models often require data to be sent to external servers, posing a compliance challenge for these sectors. Pokee-Isaac 28B addresses this by enabling long-horizon agents to accumulate and process context without data exfiltration.

The Pokee research team claims that Pokee-Isaac 28B achieves 93.3% performance on the RULER benchmark at its full 10 million token context window. They also state that the model demonstrates parity with leading cost-optimized cloud baselines on agentic benchmarks. Furthermore, the model is designed for efficient serving, capable of running on a single GPU. While the company advertises deployment on consumer-grade GPUs like the RTX 4090, the research team's reported measurements were conducted on a single B200-class GPU, suggesting the consumer-GPU claim should be considered vendor guidance.

Pokee AI offers Pokee-Isaac 28B through an OpenAI-compatible developer API and licenses it for deployment within a Virtual Private Cloud (VPC), on-premises infrastructure, or directly on devices. The launch announcement highlights Day-0 support for vLLM and SGLang, popular frameworks for efficient large language model serving. This model is particularly suited for organizations that already possess their own inference infrastructure, including mid-size and enterprise teams with dedicated platform groups, as well as device original equipment manufacturers (OEMs). Individual users without on-premise hardware are advised to utilize the hosted API, as the benefits of the "customer boundary" deployment are most realized when an existing boundary is in place.

The target industries for Pokee-Isaac 28B include healthcare and payors, financial services and insurance, defense and public sector, legal and e-discovery, and pharmaceutical or semiconductor research and development. The common requirement across these sectors is a regulatory mandate preventing data from crossing an external API boundary, rather than solely a preference for privacy. Potential applications span a wide range of complex tasks, such as comprehensive whole-repository code review, in-depth analysis of multi-year contracts and insurance claims, forensic investigations over extensive log archives, and the development of long-running tool agents capable of sustained operation and context management.