Pokee AI releases Pokee-Isaac 28B, an agentic model with 10 million token context for on-premise deployment

Pokee AI has released Pokee-Isaac 28B, a 28-billion-parameter foundation model designed for environments where data cannot leave organizational boundaries. The model features a 10-million-token context window, enabling long-horizon agentic tasks while maintaining regulated compliance requirements for on-device, on-premises, or VPC-based deployments. Long-context agents typically accumulate context faster than they resolve tasks, with every tool output and reasoning step remaining in the window. Previously, this capability was available primarily through cloud endpoints. Pokee-Isaac addresses this by delivering comparable performance to cloud baselines while running entirely within customer infrastructure. The model demonstrates strong performance across multiple benchmarks. On RULER, Pokee-Isaac maintains 93.3% accuracy at 10 million tokens, with context capabilities that exceed both GPT-5.6 Luna and Gemini 3.5 Flash Lite, which overflow at 1 million. On needle-in-haystack tests (MRCR v2) and other agentic benchmarks, the model reaches parity with cost-optimized cloud alternatives. On Terminal-Bench 2.1, it resolves 65.1% of text-compatible tasks, with red-teaming results showing the lowest direct and indirect attack success rates among tested models. Pokee AI serves Pokee-Isaac through an OpenAI-compatible API and licenses it for deployment inside a VPC, on-premises, or on-device. The company offers single-GPU serving from an RTX 4090 or equivalent, with pricing at $0.15 per million input tokens and $1.00 per million output tokens. The model has been tested on Intel Arc Pro B70, Core Ultra Series 3, and Qualcomm Snapdragon X2 Elite, demonstrating on-device deployment viability.