Qwen 3.6, ik_llama hit 50+ t/s

The pairing of Qwen 3.6 with the ik_llama fork is a watershed moment for local inference, proving that frontier-level speeds are achievable without enterprise-grade hardware. ik_llama's specialized CUDA kernel fusing provides a 26x boost in prompt processing, critical for the massive 200k+ context windows supported by the 3.6 series. The Qwen 3.6-35B-A3B MoE architecture offers a "sweet spot" for local users, fitting into consumer GPUs while rivaling Claude 3.5 Sonnet in coding and tool-calling benchmarks. This release successfully addresses the "reasoning loop" issues of the 3.5 series, where models would generate thousands of redundant tokens for simple logic tasks. Support for advanced quantization formats like UD_Q_4_K_M ensures high perplexity retention even at lower bit-widths, maximizing the utility of limited local memory.