|

Pokee AI Releases Pokee-Isaac 28B: A 10M-Token Context Agentic Model Built to Run Inside the Customer Boundary

ByRicardo

Long-horizon brokers accumulate context quicker than they resolve duties. Every device output, remark, and intermediate reasoning step stays in the window, and the two capabilities that matter — holding that context and staying coherent throughout it — have thus far been accessible virtually completely from cloud endpoints. That excludes regulated industries, public-sector establishments, and on-device purposes, the place the knowledge will not be permitted to depart the boundary in any respect. Pokee AI launched Pokee-Isaac 28B, a 28B text-only basis mannequin with a 10M-token context window, designed to run inside that boundary. The Pokee analysis workforce claims 93.3% on RULER at 10M tokens, parity with the strongest cost-optimized cloud baselines on agentic benchmarks, and a serving profile that matches a single GPU.

Is it deployable

Yes — however licensed, not open-weight. Pokee AI serves Isaac by means of an OpenAI-compatible developer API, and licenses it for deployment inside a VPC, on-premises, or on-device. The launch announcement advertises Day-0 assist for vLLM and SGLang, and single-GPU serving ranging from an RTX 4090 or equal. The analysis workforce publishes measurements solely from a single B200-class GPU, so deal with the consumer-GPU declare as vendor steering moderately than a reported end result.

  • Company stage: This matches organizations that already personal their inference stack — mid-size and enterprise groups with a platform group, plus machine OEMs. A solo practitioner with out on-prem {hardware} ought to use the hosted API as an alternative; the boundary argument solely pays off when you have a boundary.
  • Industries: Healthcare and payors, monetary companies and insurance coverage, protection and public sector, authorized and e-discovery, and pharma or semiconductor R&D. The frequent trait is a rule that the knowledge can’t cross an exterior API boundary, not a choice for privateness.
  • Applications: Whole-repository code evaluate, multi-year contract and claims evaluation, incident forensics over full log archives, and long-running device brokers that by no means want summarization or context pruning. The analysis paper makes this second level explicitly: when sufficient usable context is accessible in-boundary, reminiscence hierarchies and compression change into non-compulsory moderately than required.