I wanted a machine that could run real local models for agent work without depending on a cloud API for every call — partly cost, partly latency, and partly just wanting to understand what current consumer hardware can actually do for inference rather than training.
The build centers on an Intel Core Ultra 9 285K, an RTX 5070 Ti with 16GB of VRAM, and 128GB of DDR5. The 128GB is the deliberate choice in that list — for long-context, autonomous-agent-style workloads, system RAM for context and offload matters as much as raw GPU throughput, and 128GB gives a lot of headroom before context length becomes the limiting factor.
Ollama is the serving layer day to day, with LM Studio in the mix for quick model comparisons and for cases where a GUI is faster than a CLI for testing a prompt. The Qwen family has been the main model line running locally, since it gives a good balance of capability and VRAM footprint for a 16GB card.
The open question I am still working through is dual-GPU inference — evaluating a second RTX 5060 Ti to get roughly 32GB of combined VRAM across two cards rather than upgrading to a single larger card. I validated that the second PCIe 4.0 x4 slot on the board is usable for inference, which was the prerequisite before spending money on the second GPU.
Notes for next time: for local LLM hosts, the VRAM-vs-system-RAM tradeoff matters more than people expect going in. A 16GB card with 128GB of system RAM behaves differently for long-context work than a 16GB card with 32GB of system RAM, even though the GPU spec sheet looks identical.