Quad RTX 5060 Ti bench Qwen3.6-27B code gen at $2K

A user benchmarks a quad NVIDIA RTX 5060 Ti setup for running the Qwen3.6-27B model, targeting cost-effective code generation with high context window support.

The configuration runs the model in Q8 quantization with FP16 key-value cache and Multi-Token Prediction (MTP) enabled to maximize accuracy and speed.

Testing on a Vast AI instance yielded 608 tokens/s for cold prefill and 52.2 tokens/s for decoding at a 256K context length. The author compares this against alternatives like dual RTX 3090s and Mac M5 Max, noting the quad 5060 Ti offers better value and performance for dense models.

This setup provides excellent inference speeds for state-of-the-art code generation within a $2,000 hardware budget.