The First 10-Million-Token AI Model Built to Keep Patient Data Behind Hospital Walls

Pokee AI has released Pokee-Isaac 28B, a 28-billion-parameter foundation model designed to address a persistent barrier to clinical AI adoption: analyzing millions of tokens of patient data without violating HIPAA or transmitting protected health information to external cloud APIs. The model features a 10-million-token context window and is engineered to run within customer data boundaries—inside VPCs, on-premises data centers, or on edge devices—addressing compliance concerns that have prevented healthcare organizations from deploying long-context AI for clinical workflows.

A 10-Million-Token Window That Stays Inside the Firewall

According to the release, Pokee-Isaac 28B achieves 93.3% accuracy on the RULER benchmark at 10 million tokens, a measure of long-context retrieval performance. Competing models including GPT-5.6 Luna and Gemini 3.5 Flash Lite reportedly fail beyond one to two million tokens. On the MRCR v2 multi-needle retrieval benchmark, the model scores 0.607 at 256,000 tokens, 0.743 at 512,000 tokens, and 0.500 at one million tokens, outperforming Gemini by margins of 0.133 to 0.295 across those context lengths.

The model runs on a single B200-class GPU, with time-to-first-token of 23.6 seconds at one million tokens and 72.9 seconds at 10 million tokens. Prefill throughput ranges from 42,400 to 137,200 tokens per second, with decode speed around 335 tokens per second. Pokee AI claims the model is also compatible with consumer-grade hardware including the RTX 4090, Intel Arc Pro B70, Core Ultra Series 3, and Qualcomm Snapdragon X2 Elite processors.

For healthcare providers, the 10-million-token capacity enables processing of a decade of patient notes, lab results, imaging reports, and medication histories for a single individual—often exceeding 500,000 tokens—without lossy summarization that risks omitting critical clinical details. The model is licensed for deployment via an OpenAI-compatible API, with day-zero support for vLLM and SGLang inference frameworks. Model weights are not publicly released. Provisional pricing is set at $0.15 per million input tokens and $1.00 per million output tokens.

Agentic Benchmarks and Security Red-Teaming

On agentic task benchmarks, Pokee-Isaac 28B scores 70.94 on BFCL v4 function-calling, compared to Luna's 70.61. On τ³-bench, which measures multi-domain reasoning, it averages 0.662 across four domains versus Gemini's 0.631. On MCP-Atlas, a benchmark for multi-step task coverage, the model achieves 74.59% coverage in 9.10 turns, compared to Gemini's 14.99 turns. On Terminal-Bench 2.1, it resolves 56 of 86 tasks (65.1%), trailing Luna's 60.

Security testing using the DTAP red-teaming framework shows 36.0% direct attack success, 35.2% indirect attack success, and 35.6% combined attack success—the lowest rates in the comparison set—with 82.5% success on benign tasks. These figures suggest resistance to adversarial manipulation, a concern for high-stakes applications like fraud detection in claims databases or regulatory submission review.

Target Use Cases in Healthcare and Beyond

Pokee AI explicitly names healthcare providers, payers, pharma, financial services, defense, public sector, legal e-discovery, and semiconductor R&D as target industries. In healthcare, proposed use cases include whole-EHR reasoning for longitudinal patient record analysis, autonomous prior authorization bots that read full policy documents and patient charts, clinical coding assistants that review complete operative notes and discharge summaries, and pharmacovigilance systems that correlate adverse events across years of medication administration records.

For payers and life sciences companies, the model's on-premises deployment and long-context capabilities position it for real-world evidence extraction from longitudinal patient cohorts, claims adjudication workflows, and multi-year adverse event investigation—domains where data sovereignty and compliance requirements have historically blocked cloud-based AI adoption.

Sources

1. Pokee AI Releases Pokee-Isaac 28B: A 10M-Token Context Agentic Model Built to Run Inside the Customer Boundary — MarkTechPost