A user reports successfully running the ternary-quantized Qwen3.6 27B model on two NVIDIA RTX 3090 GPUs. The setup achieves a throughput of 60 tokens per second with stable tool calling capabilities.

  • Each GPU is allocated 100k KV cache slots.
  • Total VRAM consumption is approximately 21GB across both cards.
  • The user notes that the output quality appears good, though no coding tests were performed.