# Example configurations Start with `bin/basalt serve --config basalt_config.json`. Model paths are relative to the file. Settings: [configuration](index.md). ## One GPU (32 GB) and RAM ```json { "version": 2, "models": [{ "model": "model.basalt", "gpu": "0", "max-context": 262144, "kv": "int8" }], "engine": { "resident-experts": true, "vram-reserve-mib": 700 }, "server": { "host": "127.0.0.1", "port": 8080 } } ``` ## A GPU and a smaller second card (peer expert tier) Like the first, with `"gpu": "0,1"` and `"vision": 1`, which runs the image encoder on the second card too. ## A low-VRAM card Like the first, with `"max-context": 65536`, `"kv": "q4_0"` and `"vram-reserve-mib": 1024`. Several models: [models and the router](models.md). `"relaxed-drafting": 0.5` in an entry trades exactness under sampling for speed ([precision](../precision/qwen38-flash-next.md#relaxed-drafting)).