Pokee's 28B model holds 10 million tokens on a single GPU — and stays inside your firewall

Pokee AI just shipped a model that solves a problem most enterprises are too busy working around to admit they have. Pokee-Isaac 28B is a 28-billion-parameter text model with a native 10-million-token context window, and it runs entirely inside a VPC, on-premises server, or even on-device. No data leaves the boundary. For regulated teams in healthcare, finance, defense, and legal — sectors where “API call” is a non-starter — that alone is the headline.

The research team didn’t just claim long context; they published measurements. Isaac scores 93.3% on RULER at the full 10 million tokens. Every other model in their comparison panel, including GPT-5.6 Luna and Gemini 3.5 Flash Lite, flatlines to zero beyond 2 million tokens. On the multi-needle retrieval benchmark MRCR v2, Isaac opens a 0.295-point lead over Gemini at 1 million tokens. Those are the kind of gaps that make memory hierarchies and chunking strategies look like yesterday’s architecture.

Agentic performance is competitive without being dominant — and the paper is honest about where it loses. Isaac edges Luna on BFCL v4 (70.94 vs. 70.61) and leads the τ³-bench average at 0.662 across four domains. But on Terminal-Bench 2.1, it resolves 65.1% of text tasks to Luna’s 69.8%. That’s the one benchmark where cloud still wins, and the report states it plainly. On security, Isaac posts the lowest combined attack success rate on the DTAP red-teaming suite at 35.6%, while maintaining 82.5% benign task success.

What’s genuinely surprising is the serving profile. On a single B200-class GPU, time-to-first-token at 10 million tokens is 72.9 seconds — not fast in absolute terms, but roughly three times the TTFT at 1 million tokens despite a tenfold increase in prompt length. Prefill throughput climbs to 137,200 tokens/s at that scale. The company also claims single-GPU serving on consumer hardware starting from an RTX 4090, though published measurements only cover the B200. Pricing is provisional at $0.15 per million input tokens and $1.00 per million output. The catch is licensing: weights are not open, and deployment requires a commercial agreement. For teams that already run their own inference stack, that’s paperwork. For everyone else, there’s a hosted API.