Qwen3.6-27B KLDs – INTs and NVFPs

Will do more, but here's a start, as you're chosing your models. Remember, USE-CASE is important:

Notice the larger size of THoTD NVFP versus the other. This is because THoTD is an NVFP4A16 versus NVFP4(A4).
- NVFP4(A4) should stay in 4bit the whole time, so if you are doing batching, NVFP4(A4) may see better performance as batching occurs
Notice that huge size increase for Cyan from INT4 to BF16-INT4.
- More food for thought. Mixed-precision is amazing, but takes more space. Is 0.02 accuracy worth losing 6GB of Context? Up to you to decide.

As more come online I will add more to the graph. The more you know, the right quant for you, you grab the first time!!