Qwen3.6 quants beat smaller VRAM bets

Bigger quants can be the better local-inference choice once you’re memory-constrained, especially on MoE models where stability and runtime behavior matter as much as raw file size.

• –The smaller IQ4_XS variant hit looping issues during thinking, while the larger Q4_K_XL reportedly ran faster and more reliably
• –Throughput stayed strong even at long context, which suggests the real bottleneck is not just model size but how the quant interacts with the runtime and memory system
• –On hybrid CPU/GPU setups, a slightly larger quant can reduce pathological behavior and still improve end-to-end latency
• –Q5_K_S looks like the pragmatic pick here: close to the faster Q4 in speed, with better quality and more predictable outputs
• –This is a useful reminder for local LLM users to benchmark beyond “fits in VRAM” and test actual tokens/sec plus output stability