让模型在思维链里用标签声明"自己要看哪里",引擎解析声明并跳过大部分 KV cache 读取。零训练零样本:注意力读取降 52.0% / 31.1%,精度仅降 1.27pp / 2.75pp,解码墙钟 0.71× / 0.77×。 The model declares where to look with tags in its chain-of-thought; the engine parses the declarations and skips most KV-cache reads. Zero training, zero-shot: attention reads down 52.0% / 31.1%, accuracy down only 1.27pp / 2.75pp, decode wall-time 0.71× / 0.77×.
把个人电脑变成统一的弹性推理平台:带宽自适应执行让 8GB 笔记本跑 35B、台式机跑 284B、单张工作站 GPU 跑 753B GLM-5.2,快到能支撑真实 agent 负载。 Turns the consumer PC into one elastic inference platform: bandwidth-adaptive execution puts a 35B model on an 8GB laptop, 284B on a desktop, and 753B GLM-5.2 on a single workstation GPU — fast enough for real agent workloads.
NVFP4 低精度推理在延迟敏感场景的精度救星:步骤感知温度缩放按解码步骤动态调整 softmax 温度,不牺牲延迟地恢复量化精度。 An accuracy rescue for latency-critical NVFP4 inference: step-aware temperature scaling adapts the softmax temperature per reasoning step — recovering quantized accuracy at zero weight cost, plus a CUDA-core small-M decode kernel.
KV cache 驱逐不需要打分:保住 prompt + 每个 KV 头内均匀随机驱逐,4 模型 × 6 任务追平最强基线、60 格中 31 格显著领先,vLLM 吞吐再快 32–43%——一篇把"选择信号几乎不贡献精度"钉死的机制论文。 KV cache eviction needs no score: pin the prompt and evict uniformly at random within each head — matching the strongest evictor across 4 models × 6 reasoning tasks (significantly ahead in 31 of 60 cells) while serving 32–43% higher throughput in vLLM, with a full mechanistic account of why the selection signal buys almost nothing.
DeepSeek-V4.1-Flash 技术报告:552B 多模态 MoE 用 Causal Encoder-Decoder 让 prefill 只激活 8B;CSA2 跨层 KV 复用 + FP4 量化把全局 KV Cache 压到每 token 890 字节(V4-Flash 的 1/4、V1 的 1/437),SWA Bounded Replay 再把持久缓存压到上代 1/8——性能反超更大的 V4-Flash:DeepSWE v1.1 74.2、Codeforces 3471。 The DeepSeek-V4.1-Flash technical report: a 552B multimodal MoE whose Causal Encoder-Decoder activates just 8B in prefill; CSA2 cross-layer KV reuse + FP4 squeeze the global KV cache to 890 bytes/token (1/4 of V4-Flash, 1/437 of V1) while SWA Bounded Replay shrinks the persistent cache to 1/8 of the last generation — outperforming the larger V4-Flash: DeepSWE v1.1 74.2, Codeforces 3471.
首个专门评测多模态嵌入模型长上下文能力的基准:4 个检索任务、8 个数据集、8,460 个查询、11 个模型。核心发现:现役模型靠表面特征匹配"作弊",细粒度信息保持随长度衰退——"窗口大"≠"读得懂"。 The first benchmark dedicated to MEM long-context ability: 4 retrieval tasks, 8 datasets, 8,460 queries, 11 models. Key finding: current models "cheat" via surface-feature matching and fine-grained retention decays with length — a wide window ≠ real comprehension.