Six systems. No switch.
Everything measured.

• Six GB10 systems — NVIDIA DGX Spark and ASUS GX10 machines, 128 GB of unified memory each.
• Switchless rings. Nodes are cabled directly over their ConnectX-7 links in 2- and 4-node rings — collective traffic rides the NICs' own hardware forwarding, so every node can reach every other without putting a switch in the model-traffic path.
• Multi-node tensor-parallel serving for frontier open-weight models, quantized to NVFP4 where it makes sense.
• Everything documented. Each configuration ships with a build log, a benchmark record, and a reproducible deployment kit.

GLM-5.3-Flash

NVFP4 · TP2

Tensor-parallel serving on a DGX Spark pair, with the deployment kit published alongside its measured results.

Qwen3.8-Flash-Next

NVFP4 · TP1 / TP2

Single-Spark and pair deployments — including a hybrid checkpoint whose mixed precision beats the stock release on GB10.

DeepSeek-V4.1-Flash

TP4 · SWITCHLESS RING

Served across a four-node ring (2× DGX Spark + 2× ASUS GX10) with full benchmark documentation.

• Build. We stand the system up on our own hardware first.
• Measure. Every deployment gets a benchmark record — decode rates, prefill, concurrency — not vibes.
• Publish. Kits, tools, checkpoints, and field notes go out in public. If it doesn't hold up in the open, it doesn't ship.