Bjorn Hunter Professional Corporation
Custom, self-hosted AI engineering — Edmonton, Canada
A private three-node NVIDIA DGX Spark supercluster, built and operated end to end: hardware, fabric, serving stack, evaluation, and the local agents that run on it.
The cluster
| Compute | 3 × NVIDIA DGX Spark (GB10 Grace-Blackwell), 128 GB unified memory each |
| Fabric | direct 200 Gb/s QSFP per node pair — full triangle, no switch |
| Flagship inference | DeepSeek-V4.1-Flash-1017, EXL3, tensor-parallel across 3 nodes |
| Context | 1,048,576 tokens · 4 concurrent sequences · 2.2 GiB KV cache |
| Throughput | 53.9 tok/s single stream · 137.6 tok/s aggregate, 4-stream burst |
| Long context | 443,223-token needle retrieved correctly in 87.1 s |
Alongside the flagship model the cluster runs local vision, embedding and reranking, speech synthesis, and image/video generation — each scheduled into its own window so they never contend with inference.
Published work
- dgx-spark-supercluster-evals — independent benchmark results with every raw lm-eval file: GSM8K 92.8%, MMLU 86.1%, IFEval 88.0%, WinoGrande 79.3%, ARC 71.4%, GPQA-Diamond 52.0%, TruthfulQA 50.8%.
- dgx-spark-serving-profile — the serving profile and operating notes that make 1M context work on this hardware.
How I work
Measured, not claimed. Every number here was produced on these machines under a named configuration, with the result files published alongside it — including the superseded runs and the weak spots.