A new class of chip, designed to think. Three form factors, one architecture.
Inference moves from a general-purpose chip to silicon designed specifically for it. The difference is measurable.
Measured against unaccelerated baselines on NYMPH 1.1 reference platform.
Slots into any desktop or workstation. Frees your GPU. The full power of dedicated cognitive silicon, on-prem and air-gapped.
USB-C / Thunderbolt 4 dongle for laptops. Bus-powered. Same cognitive engine, in 120 × 70 mm. Plug it in anywhere.
The first gaming console with dedicated local AI silicon. Android AOSP, HDMI, M.2 upgradeable. Coming 2026.
| Form factor | PCIe Gen3 x4 · USB-C 3.1 module |
| Processor | RK3588 · 8-core ARM Cortex-A76/A55 · 6 TOPS NPU |
| Memory | 2–8 GB dedicated LPDDR4 |
| Storage | 64 GB eMMC + optional NVMe |
| Power | Under 5 watts |
| Compatibility | Any desktop or workstation with a PCIe x4 slot |
| Software | Cognitive runtime included |
| Manufacturing | MOKO Technologies · Zhuomao · Shenzhen |
| GPUs | Embedded NPUs | NYMPH Card | |
|---|---|---|---|
| Purpose | Gaming · HPC | Locked to SoC | AI inference |
| Form factor | PCIe | Embedded | PCIe · USB · Console |
| Persistent state | No | No | Yes |
| Quality verification | No | No | CCS |
| Power draw | 46–350 W | 5–15 W | Under 5 W |
PCIe Gen3 x4 inference accelerator with 8 GB dedicated memory.
Full-height and low-profile. Fits any ATX, microATX, or SFF build.
Install, verify, and run your first model in under 10 minutes.
Full software stack. No subscription key. No telemetry.
Render, transcribe, generate, and iterate on your own hardware — with the full model resident on your machine.
Keep context across sessions, datasets, and experiments. No re-uploading. No context window amnesia.
Deploy inference on-prem. Compliant with regulated industries. No per-token pricing, no external endpoints.
First production run begins 2026. No subscription. No recurring fees.