Pokee AI releases Pokee-Isaac 28B: a 10M-token context agentic model that runs inside the customer boundary

Pokee AI released Pokee-Isaac 28B, a 28B text-only foundation model with a 10M-token context window, aimed at regulated industries and on-device applications where data cannot leave the customer boundary. The model is licensed, not open-weight, and is served through an OpenAI-compatible API or deployed in a VPC, on-premises, or on-device, with Day-0 support for vLLM and SGLang. On RULER, Isaac scores 93.3% at 10M tokens, exceeding all baselines beyond 2M context. It leads BFCL v4 with 70.94 and τ³-bench with a 0.662 average, places second on Terminal-Bench 2.1 with 65.1%, and third on MCP-Atlas with 74.59% coverage but fewer turns. On DTAP red-teaming, Isaac records the lowest combined attack success rate at 35.6% while maintaining 82.5% benign success. Efficiency results include a TTFT of 23.6s at 1M tokens and 72.9s at 10M on a single B200-class GPU, with prefill throughput reaching 137,200 tokens/s. Pricing is provisional at $0.15/$1.00 per million input/output tokens. Isaac also runs fully on-device on Intel Arc Pro B70, Core Ultra Series 3, and Qualcomm Snapdragon X2 Elite.