#DeepSeek

DeepSeek 居然把支撑他们大规模生产服务的高性能推理框架开源了：DeepInfer 这是 DeepSeek 线上真实跑 DeepSeek 服务的同源技术栈首次对外公开，本次发布聚焦离线推理，采用 Apache 2.0 许可证，主要基于 @vllm_project 构建，并借鉴了 @sgl_project 和 TensorRT-LLM。 开源地址：https://github.com/deepseek-ai/DeepInfer DeepSeek 此前已陆续开源了推理栈的“零件”：FlashMLA（MLA 注意力内核）、DeepGEMM（矩阵乘法内核）、DeepEP（MoE 通信库）、EPLB（专家负载均衡）等，但一直缺少把它们组装成完整引擎的“骨架”。DeepInfer 补上了这一环：它是这些内核的集成者，加上调度、并行、投机解码等系统层的完整实现。配合本次同时出现的 DeepSeek-V4.1-Flash 模型，构成了“模型 + 引擎”的完整开源推理方案。硬件层面同时支持 NVIDIA SM100 GPU 和华为昇腾 950 NPU，双平台支持是它的一个鲜明特征。 # 核心技术拆解 1. 内核与运行时。 集成 FlashMLA、DeepGEMM、DeepEP、TileKernels、DeepSelect，以及用 TileLang 编写的自定义内核（位于 deepinfer/model_executor/layers/tilelang_ops）。这些是吞吐和延迟的底层保障。 2. 并行策略：Attention 与 MoE 分治。 注意力部分采用 DP（数据并行）或均衡的 CP（上下文并行，处理超长序列）；MoE 部分采用 EP（专家并行）加 EPLB（专家并行负载均衡），并支持冗余专家（默认 32 个，向上取整到能被 EP 数整除）来缓解专家负载不均。这种“注意力 DP/CP + MoE EP”的组合是 DeepSeek 生产验证过的架构。 3. DSpark 投机解码。 基于 MTP（multi-token prediction）draft token 加速解码，亮点是“置信度调度验证”；按置信度决定接受多少 draft token，而非固定规则。默认最大 draft 长度 5。这是推理加速比的主要来源之一。 4. YoCo SWA 有界重放。 针对 YoCo 式解码器 + 滑动窗口注意力（SWA）的优化：prefill 时只保留并重放有界的 token 后缀（默认 140 个，需覆盖 SWA 窗口加最大回溯长度），大幅削减 decoder prefill 的计算量。仓库还专门提供了 YoCo mask/compact 等价性验证。 5. 异步调度。 CPU 上的批调度与 GPU/NPU 上的执行重叠，做到“零开销调度器”，默认开启。 6. 批不变与确定性推理。 开启 DEEPINFER_BATCH_INVARIANT=1 后，不同批次组合下结果完全一致；再配合固定 seq_seed、关闭 MTP varlen，可实现逐 token 确定性的采样输出。这对可复现评测和回归测试（仓库用逐样本精确对比的 PPL 基线来把关）很关键，目前仅限 NVIDIA 平台。 7. 结构化输出。 原生支持 regex、JSON schema、EBNF 三种约束生成，通过 SamplingParams 的 so_args 传入。

We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program. We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.

Mozilla just published a 91-page report and it says open-weight AI is now only about 4 months behind the frontier. - 8 of OpenRouter's 10 most-used models by August token volume were open-weight, and 7 were Chinese-built, while DeepSeek became the first open model to lead the platform in weekly requests. - However, the economics are almost upside down. Open models handled roughly 20% of measured OpenRouter usage but captured only about 4% of model-layer revenue in the cited 2025 window. Mozilla attributes much of that mismatch to pricing, with closed models costing roughly 6x more per call at about 90% capability parity. The comparison shows why high usage does not necessarily translate into high revenue when one class of models is dramatically cheaper. Mozilla also warns that the revenue measurement is older than its 2026 usage data, so the current revenue split may be different. - Open models are spreading faster than they reach production: 79% of surveyed developers use them, but only 51% of open-model deployments reach production versus 63% for closed models. - DeepSeek showed that a new pretraining run may no longer be necessary for a major capability jump, gaining roughly 10 index points and then another 8 through post-training passes. -Even model diversification may provide less protection than assumed: Kimi K3 and Claude Fable 5 had a 0.72 per-task failure correlation, meaning supposed backup models often fail on the same problems.

🚀 推出 DeepSeek-V4.1-Flash：更智能、更快速、更高效。 🔹 推出我们新架构家族中最小的模型，具备原生视觉理解能力。 🔹 旨在实现更强大的能力、更快的推理、更高的吞吐量，并可扩展至更大的模型。 1/6