ASUS GX10, vLLM and OpenCode: one vs two nodes

Tim tests a self-hosted coding assistant on one ASUS Ascent GX10, then links two machines to try a larger model. Each unit has 128 GB of unified memory. vLLM serves models stored on a NAS, while OpenCode gives the AI agent access to project files and a terminal. NVIDIA Sync connects his main computer to the remote workspace through VS Code.

The task is a homelab service tracker with explicit requirements and tests. The single-node Qwen 3.6 35B FP8 run produces a working app after repeated intervention. Tim encounters stalled tool calls, broken filtering and save operations, and testing problems. He cautions that failures may come from the model, OpenCode or their interaction.

For two nodes, he uses a direct network link and a Ray cluster. Loading the larger MiniMax model requires an AWQ 4-bit quantization and memory headroom for the KV cache. The transcript mostly calls it M2.7 but later labels it M2.6, leaving the exact version unclear. Tim reports roughly 41 tokens per second after tuning and about 330 W combined during one observed workload.

The larger-model run handles the app more successfully, though it still needs supervision. Since both the model and hardware change, this is a workflow comparison rather than an isolated hardware benchmark. Tim finds local AI useful for background development work, but stresses the effort required to configure and debug the stack.