Long-context AI has largely been a cloud game, and that has left regulated industries on the sidelines. Pokee AI thinks it has the answer with Pokee-Isaac 28B, a licensed text model whose 10-million-token window is designed to run inside a customer’s own VPC, on-premises hardware, or devices, with no data crossing an external API boundary.
The company reports 93.3 percent on the RULER benchmark at the full 10 million tokens and claims parity with the strongest cost-optimized cloud baselines on agentic evaluations. Serving needs a single GPU, with day-zero support for vLLM and SGLang, and Pokee says an RTX 4090-class card can get started, though its published measurements come from one B200. The model ingests longer prompts faster, not slower: aggregate prefill speeds rise from about 42,400 tokens per second at 1 million tokens of context to 137,200 at 10 million, while decode stays near 335 tokens per second. On its agentic scorecard Isaac’s 70.94 sits barely ahead of GPT-5.6 Luna’s 70.61, a lead the technical report describes as a tie.
Healthcare payors, financial services, defense, legal, and pharma R&D are the target buyers, and the use cases run long: full-repository code review, multi-year contract analysis, and incident forensics over entire log archives. Pokee’s argument is that once the full context lives inside the boundary, teams no longer need memory hierarchies or compression schemes to cope. The trade-off is the license itself, since Isaac is not open-weight.
