Recipe

The service starts the server with the pinned recipe: the model as its first argument, --model-id bonsai2, --host 127.0.0.1 --port 18187 --api-key, --max-context and --kv-capacity at the sized window, --kv-dtype rk8v4 or rk4v4, --gdn-state-fp16, --spec dflash2 --draft-tokens 5, --vision --vision-residency overlay --vision-max-merged 12288, --max-concurrency 1, --no-webui, --host-cache-mib at the size Synoec computed for the prefix cache, and --model-suspend --sleep-idle-seconds 300; the GPU by CUDA_VISIBLE_DEVICES. The recipe is in apps/desktop/native/guardian/src/bin/synoec-local-ai.rs, the manifest in apps/desktop/assets/local-ai/ninfer.json, and the catalog facts the provider states for the model in apps/desktop/src/shared/local-ai.ts. Runtime and weights are Apache-2.0 (licensing).