One thing worth mentioning is that these request distributions will look different depending on which harness is being used, since different amounts/types of context are injected
harness 影响的不止跑分，还一路传到硬件账单：注入多少上下文、怎么切子代理，直接决定 KV cache 复用率和推理成本结构。也要看到代价——这份数据集全部是 Claude Code 轨迹，代表性靠“世界上很大一部分 agent 编码流量走 Claude Code”这个假设撑着，harness 生态一变就得重采。
Z.ai further reduced the LLM’s hardware footprint using a method called linear attention. Usually, doubling the size of a prompt quadruples the amount of memory that a model’s attention mechanism consumes.
本周唯一一条正面证据：线性注意力确实把 KV 内存从平方降到线性。但它只作用于长上下文的 KV 缓存，权重本身照样要进 HBM。所以"架构创新会削弱 HBM 依赖"依然没被证实，最多算延缓。
the KV cache has emerged as a primary performance and cost bottleneck because its size grows rapidly with context length and batch size. Limited GPU memory forces frequent recomputation, cache eviction, or spilling to storage
This precisely quantifies why longer context windows are expensive beyond just model size: KV cache grows quadratically with context, and current GPU memory can't keep pace. Each eviction or recomputation directly inflates cost-per-token — making KV cache the hidden tax on long-context AI workloads.
KV Cache 内存占用降低 10.7 倍
令人惊讶的是：KV Cache内存占用降低了惊人的10.7倍，这一数字远超普通技术优化的幅度。KV Cache是大模型推理中的主要内存消耗部分，如此大幅度的减少意味着同样的硬件可以处理更长的上下文，或者同时运行更多模型实例。
the compu-tation cost of KV caches is quadratic to the input sequence length