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

What it does

Pokee AI launched Pokee-Isaac 28B, a 28 billion parameter language model designed exclusively for text with an exceptionally large 10 million token context window. The model is built to run inside the customer’s secure environment, on-premises or within a private virtual cloud. It achieves a high score of 93.3% on the RULER benchmark at the full 10 million token length, outpacing all other tested models, which fail beyond 2 million tokens. It also leads the BFCL v4 benchmark with a 70.94 score and ranks second on Terminal-Bench 2.1. Operational efficiency includes prefill speeds of 137,200 tokens per second on a single B200 GPU and decoding rates near 335 tokens per second at full context.

Why it matters

This model pushes the boundaries of context window size far beyond current mainstream language models, which typically max out at a few thousand tokens. A 10 million token context enables more extensive reasoning, longer document understanding, and persistent memory-like capabilities in applications. Running the model inside a customer’s boundary instead of relying on external cloud deployments gives businesses greater control over privacy, compliance, and data sovereignty. The speed metrics suggest the model remains practical for real-time or near-real-time use despite its size and context length, opening opportunities for large-scale, context-heavy AI tasks in regulated industries and sensitive data environments.

Who it is for

Pokee-Isaac 28B targets companies and developers handling extremely large sequences of text, such as legal firms, scientific research, or enterprises requiring extensive document analysis. Organizations concerned with data privacy and those needing models that operate within their own infrastructure will find the licensed deployment options compelling. This can reduce reliance on external APIs, mitigating risk of data leaks or compliance breaches.

The catch

Weights are not publicly available, and licensing restricts usage to customer boundaries, so open-source or broad public use is not possible. This limits adoption to enterprises who can handle on-premise or private cloud deployments. The model’s high compute demands also require specialized, high-performance hardware like the Nvidia B200 GPU. These factors could keep the model out of reach for smaller developers without significant infrastructure or licensing budgets.

What to watch next

Focus will be on how widely Pokee-Isaac 28B adoption spreads among privacy-sensitive sectors and whether other vendors respond with similarly large context window models suitable for customer boundary deployment. Watch for improvements in efficiency or further scaling of token contexts. Also track competitive positioning as demand for private, large-context LLMs grows and as hardware advances make such models more accessible.

AI Quick Briefs Editorial Desk