1.6.5

• Support Nemotron-3.5-Super-VL (fix NemotronH config parsing)
• Reduced VRAM overhead for prefill using cache quantization
• Lower-latency drafting with hot/sparse vocabulary
• Faster AVX2 and AVX512-BW inference
• Faster model loading when offloading experts
• Support Gemma4 MTP model (Gemma4AssistantForCausalLM)
• Fix tokenizer regression from 1.6.4 causing prompt caching to fail sometimes
• Many more fixes and optimizations

Full Changelog: v1.6.4...v1.6.5