Eight hours of execution, 966 model calls, and zero generation failures: the test posted by a user running DeepSeek Harness with NInfer on a single RTX PRO 6000 captures more than a successful experiment. The setup was local: a Windows client executed all shell and file operations, while a separate machine reached over LAN handled only inference with Qwen3.8-27B and a 262K-token context window. The server did nothing except inference, keeping local operations on the client side and separating control from compute.

The workload profile was heavily skewed toward long inputs: 131.2 million input tokens versus 853.3 thousand output tokens after including compaction, with roughly 160 input tokens for every output token. The median root request was 136.6K tokens, p95 was 205.9K, and the largest reached 231.2K. Unsurprisingly, the main cost was repeated context processing rather than generation. Weighted decode speed was 104.83 output tokens per second. Root time to first token had a median of 0.8 seconds but a p95 of 136 seconds; the subagent's median was 151 seconds while its median decode phase was only 5.3 seconds. Once a request reached the GPU, it flew, while queueing and prefill created most delays.

The local tool layer performed 1,421 real operations across PowerShell, reads, edits, writes, and searches, with 30 failures and a 2.11% error rate. File tools mostly finished in milliseconds. In an agentic workload, tool reliability can be more constraining than model quality. Here the model had no generation failures, while the tools absorbed nearly all problems, shifting attention from the LLM to the pipeline.

Priced through APIs, the same workload would have cost $18.61 with DeepSeek V4 Flash, $27.26 with GPT-5.6 Luna, $270.93 with Claude Sonnet 5, and $677.32 with Claude Opus 4.6, according to the posted comparison using current rates and no cache discounts. These figures are approximate, but the pattern matters: with repeated six-figure prompts and 160 input tokens per output token, API cost scales with context length, whereas local marginal cost, after hardware, is tied mainly to energy and maintenance.

The mixed Q4/Q5/Q6 quantization helps make the run feasible on a single workstation GPU by reducing VRAM and bandwidth pressure, though it introduces precision trade-offs that may not show up in shorter benchmarks. The next planned NVFP4 profile could shift the profile further. The structural point, however, is the split between controller and inference engine over LAN. It allows file and shell operations to stay away from the inference process, a useful constraint for keeping operational data inside the local network. It does not remove management complexity, but it makes self-hosted long-horizon agentic workloads more credible.

For teams evaluating on-premise deployment, this is not a universal proof: a single GPU, a Windows client, and a LAN are not an enterprise cluster, and hardware, energy, and skills costs remain. But it signals that a portion of workloads currently paid per token can be absorbed by local machines when context is large and repetitive. The incentives shift from prompt optimization to reduce tokens, toward queue and prefill optimization to saturate the GPU. That is an architecture question more than a pricing one. AI-RADAR offers analytical frameworks on /llm-onpremise to weigh these trade-offs.