Qwen3.6-27B cache quants favor Unsloth

This is a useful local-inference benchmark, not a release story: once context gets huge, KV-cache efficiency can matter as much as weight quantization on consumer GPUs.

• –The 98,304-token context setup puts real pressure on memory, so this test is mainly about end-to-end efficiency rather than model quality in the abstract.
• –Unsloth’s result suggests its quantization path is preserving quality better than the alternatives in this specific AMD setup.
• –If q8_0 really stays flat on perplexity, that is usually the safest default for people who care more about stability than squeezing every last byte.
• –q5_0 and q5_1 tend to live in a awkward middle zone: enough compression to matter, but not always enough extra upside to become the obvious recommendation.
• –The practical lesson is to benchmark the full stack, including cache format and context length, instead of only comparing GGUF file sizes.