In a significant shift for the generative artificial intelligence landscape, Pokee AI has officially unveiled Pokee-Isaac 28B, a text-only foundation model designed specifically to address the "context overflow" challenges inherent in long-horizon agentic workflows. Boasting a massive 10-million-token context window, the 28-billion-parameter model aims to bridge the gap between high-performance cloud-based LLMs and the stringent security requirements of regulated industries. Unlike its primary competitors from Google and OpenAI, which largely operate via cloud endpoints, Isaac 28B is engineered for deployment within private virtual clouds (VPC), on-premises servers, and high-end edge devices.
The release comes at a pivotal moment in AI development, as "agentic" systems—models capable of using tools, browsing the web, and executing multi-step reasoning—become the industry standard. These systems accumulate context at an exponential rate; every observation, tool output, and intermediate thought remains within the model’s "working memory." For complex tasks, such as analyzing thousands of legal documents or managing month-long software engineering projects, standard context windows of 128,000 or even 1 million tokens often prove insufficient. Pokee AI’s 10-million-token capability represents a tenfold increase over many current industry leaders, potentially unlocking new use cases in defense, healthcare, and finance where data sovereignty is a non-negotiable requirement.
Technical Architecture and the Privacy Imperative
The core value proposition of Pokee-Isaac 28B lies in its "inside the boundary" deployment model. Historically, organizations in regulated sectors—such as public-sector institutions and pharmaceutical researchers—have been forced to choose between the high-capacity reasoning of cloud-hosted models (like GPT-4o or Gemini 1.5 Pro) and the security of smaller, less capable local models. Pokee AI seeks to eliminate this trade-off by offering a mid-sized model (28B) that rivals the performance of much larger cloud baselines while remaining small enough to fit on a single high-end GPU.
The model is released under a licensed, rather than open-weight, distribution strategy. This allows Pokee AI to maintain control over the intellectual property while providing enterprise customers with the necessary binaries to run the model entirely offline. The research team confirmed "Day-0" support for popular inference frameworks, including vLLM and SGLang, ensuring that DevOps teams can integrate the model into existing pipelines with minimal friction.
A Chronology of the Context Window Evolution
To understand the impact of Isaac 28B, it is necessary to look at the rapid expansion of context windows over the last three years. In early 2023, a context window of 8,000 to 16,000 tokens was considered standard. The release of GPT-4 Turbo expanded this to 128,000 tokens, which was quickly followed by Anthropic’s Claude 2.1 (200,000 tokens) and eventually Google’s Gemini 1.5 Pro, which pushed the boundary to 1 million and later 2 million tokens.
The progression has been driven by the need for "Long-Context Retrieval" (LCR). However, as context windows grew, the "Lost in the Middle" phenomenon became a documented weakness, where models would lose track of information placed in the center of a long prompt. Pokee AI’s jump to 10 million tokens is not merely a quantitative increase but a qualitative shift toward "infinite-feeling" memory for digital agents. This chronology highlights a move away from "retrieval-augmented generation" (RAG) as a crutch and toward models that can hold entire codebases or historical archives directly in their active attention mechanism.
Benchmark Performance: Retrieval and Reasoning
Pokee AI provided extensive benchmarking data to validate Isaac 28B’s performance against industry titans. On the RULER (Retrieval from Ultra-Long Sequences) benchmark, which tests a model’s ability to extract specific information from increasingly large datasets, Isaac 28B maintained a 93.3% accuracy rate across all tested lengths, including the maximum 10-million-token tier. In comparison, GPT-5.6 Luna and Gemini 3.5 Flash Lite—two notable cost-optimized cloud models—showed parity up to the 512,000-token mark but suffered from context overflow and performance degradation at 1 million tokens and beyond.
Further testing on the Multi-Needle Context Retrieval (MRCR) v2 benchmark showcased the model’s precision. When tasked with finding eight specific "needles" (pieces of information) hidden within a "haystack" of noise, Isaac 28B scored 0.607 at 256,000 tokens and 0.743 at 512,000 tokens. Even at 1 million tokens, it maintained a score of 0.500. Analysts noted that the performance margin between Isaac and Google’s Gemini widened significantly as the context length increased, growing from a 0.133 lead to a 0.295 lead as the volume of data doubled.
In agent-specific evaluations, Isaac 28B demonstrated a high degree of "tool-call" proficiency. On the Berkeley Function Calling Leaderboard (BFCL) v4, it achieved a score of 70.94, placing it in a statistical tie with GPT-5.6 Luna (70.61). On the $tau^3$-bench, which measures performance across four domains including banking and retail, Isaac averaged 0.662, surpassing Gemini’s 0.631.
One of the more telling metrics was found in the MCP-Atlas (Model Context Protocol) benchmark. While Isaac placed third in overall coverage at 74.59%, it demonstrated superior efficiency, resolving tasks in an average of 9.10 turns. This was significantly faster than Gemini, which required 14.99 turns to achieve similar results. This efficiency suggests that Isaac is better at "zeroing in" on the correct tool or reasoning path, reducing the computational cost and time-to-resolution for the end-user.
Hardware Requirements and Serving Profiles
One of the most ambitious claims made by the Pokee research team is the model’s ability to run on a single GPU. While the technical report primarily features measurements from NVIDIA B200-class GPUs (the latest Blackwell architecture), the company advertises that Isaac 28B can be served on consumer-grade hardware, specifically the RTX 4090 with 24GB of VRAM, likely utilizing advanced quantization techniques.
For enterprise-grade deployments, the serving profile on a B200 is impressive. The Time to First Token (TTFT) was recorded at 23.6 seconds for a 1-million-token prompt and 72.9 seconds for a 10-million-token prompt. While a 72-second wait may seem long for a chatbot, for an agent processing millions of words of documentation, it represents a significant speedup over manual human review or multi-stage RAG systems. Furthermore, the prefill throughput scales efficiently with context; it rises from 42,400 tokens per second at lower context levels to 137,200 tokens per second at the 10-million-token mark. This means that a ten-fold increase in the prompt size only results in a three-fold increase in waiting time.
Beyond the data center, Pokee AI is targeting the burgeoning "AI PC" market. Isaac 28B is optimized to run on Intel Arc Pro B70 GPUs and the upcoming Intel Core Ultra Series 3 (Panther Lake) processors. It also supports the Qualcomm Snapdragon X2 Elite, signaling a future where massive-context AI agents can live locally on a professional’s laptop, operating without an internet connection.
Security, Red-Teaming, and Safety
Given its target market in regulated industries, Pokee AI subjected Isaac 28B to rigorous red-teaming via the DTAP (Dynamic Threat Assessment Protocol). The model recorded the lowest attack success rates in its class, with a combined score of 35.6% (lower is better in this context). This includes resistance to both direct prompt injection and indirect attacks where malicious instructions are hidden within a long-context document.
The report did note a slight discrepancy in testing conditions: while baseline models were tested using stock runners, Isaac was tested using the proprietary Pokee harness, which may include additional safety filtering at the inference level. Nevertheless, a benign success rate of 82.5% suggests that these safety guardrails do not significantly impede the model’s utility for legitimate tasks.
Economic Implications and Industry Reaction
The pricing for Isaac 28B is currently set at a provisional list price of $0.15 per million input tokens and $1.00 per million output tokens. This positioning is highly aggressive, aiming to undercut the operational costs of cloud providers while offering the added value of local deployment. By providing a "predictable cost" model for VPC-hosted instances, Pokee AI is appealing to Chief Information Officers (CIOs) who are wary of the variable and often opaque billing associated with cloud LLM APIs.
Industry analysts have reacted with cautious optimism. "The ability to keep 10 million tokens of context within a single organization’s firewall is a game-changer for legal and medical sectors," said one independent AI researcher. "However, the true test will be the model’s ‘coherence’ over those 10 million tokens. It is one thing to retrieve a fact; it is another to maintain a consistent reasoning chain over a document the size of twenty novels."
The only area where Isaac 28B trailed its cloud counterparts was in the Terminal-Bench 2.1 evaluation, which tests the model’s ability to interact with a command-line interface. Isaac resolved 56 of 86 tasks (65.1%), while GPT-5.6 Luna resolved 60. The Pokee research team acknowledged this gap, stating that cloud-based models still hold a slight edge in raw "coding and terminal logic," likely due to the massive scale of their training data.
Broader Impact and Future Outlook
The launch of Pokee-Isaac 28B marks a transition in the AI industry from a "bigger is better" philosophy to one of "smarter and more private." By focusing on a 28B parameter size, Pokee AI has found a "sweet spot" that allows for sophisticated reasoning and massive context without requiring a massive server farm to operate.
As organizations move away from simple chatbots and toward autonomous agents that can manage entire workflows, the demand for long-context, secure models will only grow. If Isaac 28B performs as well in real-world enterprise environments as it has in initial benchmarks, it could force a shift in how cloud providers like Google and Microsoft package their AI offerings, potentially leading to more "edge-friendly" versions of their flagship models.
For now, Pokee AI has carved out a unique niche: providing the massive memory of a cloud giant with the privacy and control of a local installation. As the AI community continues to digest the technical report and initial performance data, the focus will turn to how Isaac 28B handles the messy, unstructured data of the real world—and whether it can truly maintain its 93.3% accuracy when the "haystack" is not a benchmark, but a decade’s worth of corporate archives.
