Pokee AI releases 28B model with a 10-million-token context

Pokee AI released Pokee-Isaac 28B, a 28-billion-parameter text-only agentic model built to run inside a customer's own infrastructure. It scores 93.3% on the RULER long-context benchmark at 10 million tokens, where every compared baseline scores 0% beyond 2 million, and leads the BFCL v4 tool-use benchmark. Weights aren't published; it's licensed for VPC, on-premises, or on-device deployment at $0.15/$1.00 per million tokens.

Why it matters: Most long-context claims degrade badly past a couple million tokens, so a benchmark showing usable recall at 10M, if it holds up independently, would be a genuine capability jump for agentic use cases like whole-codebase reasoning. The closed-weight, on-prem licensing model also signals a strategy aimed at enterprises unwilling to send data to a cloud API.