Build guide
Dual-GPU workstation tuned for agent workloads with 16K context depth.
Budget
$4,500
Profile
AGENT HOST
Target model
Qwen 3 Coder 30B-A3B (MoE)
337.5tok/s
286.9–388.2 tok/s decode on Qwen 3 Coder 30B-A3B (MoE)
Value: 75.00 tok/s per $1k
Agent pipelines are VRAM-hungry in a different way than inference: a 16K context window for a coding model like Qwen 3 Coder 30B adds several gigabytes of KV cache and tool-call history on top of the model weights. A single 24 GB card runs out of headroom quickly when you add tool call history and multi-turn context. This dual-GPU config pools VRAM to comfortably hold the model and a full 16K cache simultaneously, keeping tok/s stable even as context fills. Prefill for a 16K prompt runs on the GPU — it is compute-bound, so CPU choice doesn't move time-to-first-token (see our methodology page); the 7800X3D is here to keep the orchestration layer between model calls responsive, not to accelerate prefill. 64 GB of system RAM ensures the orchestration layer (LangChain, CrewAI, or custom tooling) doesn't compete with model memory. This build isn't optimized for raw throughput on short prompts — if you're running batch 8B workloads, the Home Inference 4090 is faster per dollar.