The best way to run the crayfish model locally is on a Windows Mini PC with a Ryzen 7 processor – because we can also run the Qwen 3 model locally! Environment: Windows 11 Pro, AMD Ryzen 7 7730U with Radeon graphics 16 cores, 32GB RAM 1) In PowerShell Admin, install scoop: Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser Invoke-RestMethod -Uri get.scoop.sh | Invoke-Expression 2) Install llmfit using scoop: scoop install llmfit 3) Run llmfit and find the most suitable open-source local large model for this computer: Qwen3.5-35B-A3B (MoE architecture, 35B parameters in total but only 3B are activated), with extremely high inference efficiency, requiring all weights to load. Q3_K_M ≈ 16-17GB (16GB RAM) Q4_K_M ≈ 21-22GB (32GB RAM) 4) Install LM Studio (Windows version) lmstudio.ai Double-click to install after downloading 5) Run lmstudio, search for Qwen3.5-35B-A3B Select: Qwen3.5-35B-A3B-GGUF version Download: Q3_K_M Q4_K_M 6) Select the following parameters to start your downloaded model: Context Length: 16384 (adjust to 32768 if you feel it's stable) GPU Offload: Increase to 40 (adjust to 35 if loading fails) CPU Thread Pool Size: Increase to 16 Evaluation Batch Size: Change to 512 Max Concurrent Predictions: Keep 4 Unified KV Cache: Keep enabled Offload KV Cache to GPU Memory: Keep enabled Number of Experts: Keep at 8 Number of layers for which to force MoE weights onto CPU: Change to 0 to prevent MoE layers from returning to the CPU, enabling full GPU acceleration Flash Attention: Enabled Keep Model in Memory: Enabled Try mmap: Enabled RoPE Auto Remember settings: Check the box for one-click loading next time Click Load Model. The first load will take 1-3 minutes (full offload requires moving large files), please be patient. After successful loading, directly test the chat. Tips: Setting the iGPU shared memory to 8GB in the BIOS can further improve speed. Start chatting! 100% localized LLM, safe and reliable.