Pokee AI launches Pokee-Isaac 28B with 10M context

Pokee AI has introduced Pokee-Isaac 28B, a text-only foundation model featuring a 10-million-token context window. Unlike typical long-context models that rely on cloud endpoints, this 28-billion-parameter model is licensed for deployment inside a customer's virtual private cloud, on-premises, or on-device. This targets regulated sectors like healthcare, finance, defense, and legal services, where data privacy rules prevent information from leaving local boundaries. The model offers Day-0 support for vLLM and SGLang, and can run on hardware ranging from an RTX 4090 up to enterprise chips, as well as on-device processors like the Intel Arc Pro B70, Core Ultra Series 3, and Qualcomm Snapdragon X2 Elite.

In evaluations, Pokee-Isaac 28B achieved a 93.3% score on the RULER benchmark at its full 10-million-token capacity, whereas competing models dropped to zero beyond 2 million tokens. On the MRCR v2 eight-needle test, the model scored 0.607 at 256K, 0.743 at 512K, and 0.500 at 1M tokens, widening its lead over Gemini. It scored 70.94 on BFCL v4, compared to GPT-5.6 Luna's 70.61. On the tau-bench, it averaged 0.662 across four domains, beating Gemini's 0.631, though both struggled on banking at 0.186. On MCP-Atlas, it placed third with 74.59% coverage but required only 9.10 turns per task compared to Gemini's 14.99. It fell slightly behind Luna on Terminal-Bench 2.1, resolving 56 of 86 tasks (65.1%) compared to Luna's 60.

For security, the model registered the lowest direct, indirect, and combined attack success rates on DTAP red-teaming at 36.0, 35.2, and 35.6 respectively, while maintaining an 82.5 benign success rate. Running on a single B200-class GPU, its time-to-first-token was 23.6 seconds at 1M tokens and 72.9 seconds at 10M tokens. Prefill throughput scaled from 42,400 to 137,200 tokens per second, while decode speeds remained steady near 335 tokens per second. Pokee AI has set provisional list pricing at $0.15 per million input tokens and $1.00 per million output tokens.

For practitioners, this release eliminates the need for complex memory hierarchies or context pruning when building long-running agents. Developers can now feed entire code repositories, multi-year contracts, or massive log archives directly into a local model without worrying about data leakage or cloud API limits.