Awq Quantization

This gives Claude the ability to apply AWQ quantization, a 4-bit compression technique that preserves important weights based on activation patterns. You're looking at roughly 3x speedup over FP16 with minimal accuracy loss, which makes it solid for deploying instruction-tuned models in production. The skill documentation is clear about when to use it: if you're running vLLM on Ampere or newer GPUs and want better generalization than GPTQ offers. Worth noting it originally came from zechenzhangagi's AI research skills collection before being templated here. The tradeoff is ecosystem compatibility, GPTQ still has broader tool support if that matters for your stack.