28B-parameter proprietary non-decoder-only architecture from Pokee AI. The model is pitched as the first real 10M-token context frontier-class agentic model that fits on a single GPU - starting from an RTX 4090 (24GB) for the quantized form.
Vendor-reported highlights. 93.3% RULER at 10M-token context; up to 137K tokens/s prefill on one B200 with a 10M-token context; leads BFCL v4 and tau^3-bench in Pokee’s evaluation; lowest combined attack success rate among evaluated models on the DTAP security red-teaming benchmark.
Deployment and pricing. Closed weights. Hosted API is $0.15/M input, $1/M output. Also deployable in your VPC, on-premises, or on-device with Day-0 support for vLLM and SGLang. Technical docs and API console live at https://console.pokee.ai.
Honest framing. The architecture and most benchmarks are vendor-reported; the 10M context and single-GPU deployment claim are the concrete differentiators to watch as independent evals appear. The 28B size is consistent with the stated RTX 4090 footprint, but treat the throughput and safety numbers as claims until replicated.