Qwen 3.8 27B built this locally on my RTX 5090 with DeepSeek Harness

I’ve been testing Qwen 3.8 27B as a long-running coding agent on my RTX 5090. I used GPT-5.6 Sol to help write a short plan md for FloodLayer, then let Qwen execute the build in DeepSeek Harness. FloodLayer is a small 3D AEC sandbox where water follows the actual floor slope, moves toward drains, pools, and can escape through thresholds. What finally worked well for me was 131K context, full GPU offload with ngl 99, Flash Attention, Q8 KV cache, parallel 1, and MTP with draft max 2. In DSH I set contextWindow to 131072 and maxTokens to 16384. Loading the model directly instead of using router/preset mode was also much more stable for me. Without MTP I was getting around 54 tok/s. With MTP I’m seeing roughly 70–100 tok/s depending on context length and draft acceptance. Auto-compaction is working now too, so it can keep going for much longer without constantly needing manual continue. Still testing the long-run behavior, but this is the first setup where local Qwen genuinely feels useful as a serious coding agent.