Qwen 3.5 hits 72K context with TurboQuant

TurboQuant's integration into llama.cpp signals a shift in local LLM optimization from weight quantization to context cache efficiency. TQ3/TQ4 compression on the Value cache provides a 4.83x multiplier, essential for fitting long-context models into limited Unified Memory while keeping perplexity increases extremely low. The combination of 27B model weights and an 8-bit Key cache hits the performance sweet spot for Mac-based developer workflows, reducing generation stutter during complex reasoning tasks.