Running Qwen3.6-35B on two used P102-100 GPUs for $100

The article demonstrates running the Qwen3.6-35B-A3B model using two NVIDIA P102-100 graphics cards, which cost approximately $100 combined. This setup provides 20GB of VRAM and a memory bandwidth of 448GB/s.

• The system supports 3 concurrent users with context windows up to 32768 tokens.
• Inference speeds reached approximately 23.5 tokens per second during testing.
• The configuration utilizes llama.cpp to manage the model loading and slot allocation across the two GPUs.

This approach offers a low-cost alternative for running large language models locally, providing better speed and context capacity than cheaper cards with less VRAM.