Edge Command Workbench

Edge Command Workbench gives a local browser a small, focused interface for choosing a machine command from a short request and a list of allowed options. The model and tokenizer run on an RK3576 device, so the request can stay on the local network.

Required hardware

The existing validated package is prepared for RK3576/RKNN. The historical rk3576_remote preset uses the independent Laya/RKNN 0.1.0 paths and remains separate from the current Kev RKLLM presets. RKLLM Kev has been device-tested on RK3576 test platform and RK3588 test platform: RK3576 deployment requests are guarded at 2048 tokens while the engine export capacity is 8192; RK3588 uses an 8192-token deployment context. Existing Jetson TensorRT deployment is maintained in the application Docker sources; this solution page does not claim native Kev support on Jetson.

Network and storage

• SSH access is required for deployment.
• Keep at least 4GB free on the target filesystem.
• Docker and /usr/lib/librknnrt.so must already be installed on the target.
• The container runs without privileged mode using the RK3576 NPU DRM nodes, read-only device-tree paths, and read-only RKNN runtime mounts.
• The first deployment downloads the pinned archive, model, and tokenizer from the configured HF endpoint. The default is https://hf-mirror.com.

Kev RKLLM artifact set

• RK3576 and RK3588 engines, the FP32 Q/K head, tokenizer, metadata, RKLLM 1.3.1 runtime, and license notice are separate pinned artifacts.
• The container loads one local model family at a time. MODEL_SERVICES_JSON can point another model family to an existing HTTP service without loading both families into RAM.
• The tested devices and context limits are listed in the support matrix; other devices with the same SoC are not assumed to have the same memory headroom or unlimited context.

Optional voice

Voice input is available only when an existing compatible local voice service is supplied. The solution does not automatically install a voice service. Browser microphone access requires HTTPS or localhost.

gallery/kev-final/ contains the Chinese and English result screenshots, the pending state, and a screenshot montage from the 2026-10-11 RK3588 test platform candidate QA. SHA256 values, service URLs, health response, and request metadata are recorded in gallery/kev-final/README.md and kev-final-evidence-20261011.json. The montage shows the text-decision flow. When voice is used, connect an existing compatible ASR service and authorize the microphone from HTTPS or localhost; the protocol evidence is in assets/evidence/voice-proxy-contract-20261011/.

Context-500 QA fixture

The reproducible context-500 fixture is archived in assets/evidence/context500-20261011/. It contains 120 original and 120 reverse cases, each assembled to 500 tokens. The frozen dataset, metadata, runner, raw JSONL, summaries, logs, and monitoring records are included for both devices.

The fixture uses generated documentation context to hold the assembled input at 500 tokens. The measurements cover this fixed generated background and candidate set; metrics for real business commands require a separately labeled test set. The workbench returns a candidate option ID for the calling system to process. The source, generator, dataset SHA, runner SHA, and device bindings are recorded with the evidence.

Real browser switch evidence is in gallery/kev-final-switch/: the Laya page chip/result both show rknn / 90 tokens, while the Kev page chip/result both show rkllm / 8192 tokens. The same English request was run after selecting each profile from the dropdown; Test and connect made a real GET /api/models request and received HTTP 200. The assets document the final UI readability rounding (score to three decimals and milliseconds to one decimal); raw API values remain in the evidence JSON.