Editorial illustration for Pokee AI's 28B Model Maintains 93.3% Accuracy at 10M Tokens
Pokee-Isaac 28B Hits 93.3% Accuracy at 10M Tokens
Pokee AI put out Pokee-Isaac 28B on Tuesday, a text-only foundation model with a 10-million-token context window built to run behind a customer's own firewall rather than through a shared cloud endpoint. That distinction matters more than the parameter count. Long-running agents pile up tool outputs, observations, and reasoning traces inside their context window faster than they finish tasks, and until now the models capable of holding that much information and staying coherent across it have lived almost exclusively on hyperscaler infrastructure. Hospitals, government agencies, and device manufacturers that can't let data leave their own systems have been locked out of that capability by default.
Pokee AI is positioning Isaac as the fix: a 28-billion-parameter model licensed for deployment inside a VPC, on-premises, or directly on-device, served through an OpenAI-compatible API. The company says it fits on a single GPU and ships with Day-0 support for vLLM and SGLang. Whether the numbers hold up outside vendor benchmarks is the real question, and the research team's own reporting leaves some gaps worth flagging before anyone signs a deployment contract.
The Pokee research team claims 93.3% on RULER at 10M tokens, parity with the strongest cost-optimized cloud baselines on agentic benchmarks, and a serving profile that fits a single GPU.
Why this matters
The RULER numbers matter less for the raw accuracy figure than for what happens after 512K, where GPT-5.6 Luna and Gemini 3.5 Flash Lite hit context-overflow and Isaac keeps going to 10M. That gap is the actual product. Cloud APIs have owned long-context agents for two years now, which meant anyone bound by data-residency rules, banks, hospitals, government agencies, was locked out of the workflows that need the most context in the first place. A 28B model that stays deployable inside a customer's own boundary changes who gets to build these systems, not just how well they perform.
We'd still want to see the MRCR v2 dip at 1M (0.500, down from 0.743 at 512K) stress-tested against real retrieval tasks, not synthetic needle benchmarks, before treating this as settled. Multi-needle recall dropping by a third at longer context is the kind of detail that gets buried in a headline number. For teams building compliance-bound agents, the question now is whether that degradation shows up in production, not in RULER.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Backflip AI converts 3D scans to CAD models in minutes
From Archives Mistral AI’s Small Model Matches Safety Scores of Models 7× Larger
Common Questions Answered
What is the key advantage of Pokee-Isaac 28B's 10-million-token context window?
Pokee-Isaac 28B's 10M token context window allows long-running agents to accumulate tool outputs, observations, and reasoning traces without losing coherence, which is critical for agentic workflows that previously required cloud-based models. This capability enables organizations with data-residency requirements, such as banks, hospitals, and government agencies, to deploy advanced AI agents on their own infrastructure rather than relying on shared cloud endpoints.
How does Pokee-Isaac 28B maintain 93.3% accuracy on RULER benchmarks at 10M tokens?
The Pokee research team achieved 93.3% accuracy on the RULER benchmark at 10M tokens while maintaining parity with leading cost-optimized cloud baselines on agentic benchmarks. This performance is particularly significant because the model can run on a single GPU, making it practical for on-premises deployment without sacrificing accuracy at extreme context lengths.
Why does Pokee-Isaac 28B's performance beyond 512K tokens matter compared to GPT-5.6 Luna and Gemini 3.5 Flash Lite?
While GPT-5.6 Luna and Gemini 3.5 Flash Lite experience context-overflow limitations after 512K tokens, Pokee-Isaac 28B continues functioning reliably all the way to 10M tokens. This extended capability gap represents a significant product advantage, as it enables organizations previously locked out of long-context agent workflows due to data-residency constraints to now access these capabilities on their own infrastructure.
What deployment advantage does running Pokee-Isaac 28B behind a customer's firewall provide?
Deploying Pokee-Isaac 28B on a customer's own infrastructure rather than through a shared cloud endpoint ensures data residency compliance for regulated industries like banking, healthcare, and government. This on-premises deployment capability, combined with the model's ability to fit on a single GPU, eliminates the previous two-year monopoly that cloud APIs held over long-context agent workflows.
Further Reading
- Pokee AI Releases Pokee-Isaac 28B: A 10M-Token Context Agentic Model Built to Run Inside the Customer Boundary - MarkTechPost
- Pokee AI launches 28B agent model, claims 10 million tokens on a single GPU - RuntimeWire
- Pokee-Isaac 28B: A Real 10M-Token Context Model on One GPU - ExplainX AI
- pokee-isaac-28b.md - The Satyajit AI
- Pokee-Isaac 28B mise sur le très long contexte - LHC Media
