Beyond TurboQuant: Together AI Brings 2-bit KV Cache to Real Serving

超越TurboQuant：Together AI把2-bit KV Cache推向真实服务

Together AI, the University of Sydney, and UIUC introduced OSCAR, a 2-bit KV Cache quantization method that uses about 2.28 effective bits per KV element and scores 71.86 on Qwen3-4B-Thinking, 40.1 points above TurboQuant.

Why it matters: HKR-H/K/R all pass: OSCAR links 2-bit KV cache to serving and provides concrete scores. The topic is still low-level inference optimization, so it lands in featured rather than same-day must-write.