DeepSeek R1 Distill Llama 8B (Q6_K) 1.77 tokens per second Gemma 3 4B (Q8) 3.08 tokens per second EXAONE Deep 7.8B (Q6_K) 1.55 tokens per second Qwen 4B (Q8_0) 2.13 tokens per second (thinking mode) 2.45 tokens per second (non-thinking mode)