There was a study specifically related to Qwen3.8 27B that showed that kv cache quantization has almost no impact on this model all the way to q4: https://arxi

There was a study specifically related to Qwen3.8 27B that showed that kv cache quantization has almost no impact on this model all the way to q4: https://arxiv.org/html/2609.04098 https://arxiv.org/html/2609.04098