RTX 5080 Finds Qwen3.6 Sweet Spot

The real constraint here is not whether a model fits, but whether it stays fast enough to work in an agent loop without becoming annoying. For this hardware, the sweet spot is a quantized 27B to 35B-class model, not a tiny 7B coder.

• –Qwen3.6-35B-A3B is explicitly aimed at agentic coding, with stronger repository-level reasoning and tool-calling behavior than older local picks.
• –Qwen2.5-Coder-32B-Instruct is still the dense-code baseline to beat if you want a more classic coding-focused model with long context.
• –4-bit quantization is the practical lane on 16GB VRAM; 64GB system RAM gives you enough headroom for partial CPU offload, but speed will drop as more layers spill out of VRAM.
• –If you care about autonomous coding workflows, prioritize instruction-following, tool use, and latency over raw parameter count.
• –The bigger takeaway: consumer GPUs are finally good enough for serious local coding agents, but only if you pick models optimized for efficiency, not just size.