Pokee AI releases 28B model with a 10-million-token context
Pokee AI has released Pokee-Isaac 28B, a text-only foundation model designed to process up to 10 million tokens while running within a customer-controlled boundary. The company says it achieved 93.3% on RULER at the 10M-token length, while every baseline in the comparison panel scored 0.0 beyond 2M tokens. The model is intended for regulated organizations and applications such as repository-wide code review, contract analysis, log forensics and long-running tool use.
Isaac is available through an OpenAI-compatible API and can be licensed for deployment in a VPC, on premises or on devices. Weights have not been published. The launch advertises support for vLLM and SGLang and single-GPU serving from an RTX 4090 or equivalent, although the reported measurements were collected on one B200-class GPU, making the consumer-GPU statement vendor guidance.
On BFCL v4, Isaac scored 70.94, and its average on τ³-bench was 0.662. It resolved 56 of 86 text-compatible Terminal-Bench 2.1 tasks, below Luna’s 60, and ranked third on MCP-Atlas with 74.59% coverage. Its combined DTAP attack success rate was 35.6%, alongside 82.5% benign task success. Provisional pricing is $0.15 per million input tokens and $1.00 per million output tokens.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.