<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0">
  <channel>
    <title>ClawSphere 企业 AI 研究院</title>
    <link>https://clawsphere.io/research/</link>
    <description>数字员工、AI Agent、基座模型与企业 AI 研究。</description>
    <language>zh-CN</language>
    <lastBuildDate>Thu, 12 Nov 2026 08:15:06 GMT</lastBuildDate>
    <item>
      <title>LFM2.5 Encoder：企业本地 RAG 与策略路由的 CPU 下沉路径</title>
      <link>https://clawsphere.io/research/lfm25-encoder-enterprise-local-rag-routing-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/lfm25-encoder-enterprise-local-rag-routing-2026/</guid>
      <pubDate>Mon, 05 Oct 2026 00:00:00 GMT</pubDate>
      <description>Liquid AI LFM2.5 Encoder 230M/350M 把编码、检索前置和策略路由带到 CPU 与边缘设备。本文核对官方模型卡与发布说明，分析上下文、运行时、许可和企业部署边界。</description>
      <category>企业实践</category>
    </item>
    <item>
      <title>Manus 2.0、Meta Muse 与字节腾讯类 Muse 产品：企业 Agent 应该买什么</title>
      <link>https://clawsphere.io/research/manus-2-muse-china-agent-enterprise-routing-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/manus-2-muse-china-agent-enterprise-routing-2026/</guid>
      <pubDate>Tue, 29 Sep 2026 00:00:00 GMT</pubDate>
      <description>对照 Manus 2.0、Meta Muse、字节 Seed2.1 与腾讯 WorkBuddy、QClaw、Hy4 的官方更新，拆解执行环境、记忆、审批、成本与企业部署边界。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>Jev 模型为什么突然走红：把企业 AI 的判断层从聊天模型中拆出来</title>
      <link>https://clawsphere.io/research/jev-system-one-model-enterprise-routing/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/jev-system-one-model-enterprise-routing/</guid>
      <pubDate>Tue, 22 Sep 2026 00:00:00 GMT</pubDate>
      <description>基于 TypeSafe AI 官方博客与文档，拆解 Jev 的 System One 设计、Choice/Score/Noul 输出、置信度边界，以及它对企业路由、客服分流和数字员工治理的真实价值。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>Kimi K3 NVFP4 私有化部署：8 张 B300 为何仍不是即装即用</title>
      <link>https://clawsphere.io/research/kimi-k3-nvfp4-enterprise-deployment-readiness/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/kimi-k3-nvfp4-enterprise-deployment-readiness/</guid>
      <pubDate>Sun, 30 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 NVIDIA、Moonshot AI、vLLM 与 SGLang 官方资料，拆解 Kimi K3 NVFP4 的 1.6TB 权重、8×B300 配方、上下文、运行时成熟度和企业采购边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>GLM-5.3 编码 Agent 安全治理：网络攻击能力增强后，企业如何部署</title>
      <link>https://clawsphere.io/research/glm-5-3-coding-agent-cyber-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/glm-5-3-coding-agent-cyber-governance/</guid>
      <pubDate>Sat, 29 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 Z.ai 官方模型卡、产品文档与许可，拆解 GLM-5.3 的 744B-A40B 架构、1M 上下文、网络安全能力、集群成本和企业编码 Agent 治理边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>UI-Venus-2-9B 企业部署研究：开放权重 GUI Agent 如何进入数字员工治理</title>
      <link>https://clawsphere.io/research/ui-venus-2-enterprise-gui-agent-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/ui-venus-2-enterprise-gui-agent-governance/</guid>
      <pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 inclusionAI 官方模型卡、代码仓库与项目页，拆解 UI-Venus-2-9B 的闭环执行架构、本地部署资源、许可冲突和企业治理边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>Qwen3.8-Flash-Next 与 GLM-5.3-Flash：Flash 模型为何仍是服务器级部署</title>
      <link>https://clawsphere.io/research/flash-model-server-deployment-qwen-glm/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/flash-model-server-deployment-qwen-glm/</guid>
      <pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 Qwen 与 Z.ai 官方模型卡，拆解两款新 Flash 多模态 MoE 的参数、上下文、许可、运行时和企业私有化边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>Granite 4.2 本地部署研究：3B 推理模型如何进入企业 Agent 路由</title>
      <link>https://clawsphere.io/research/granite-4-2-enterprise-agent-local-deployment/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/granite-4-2-enterprise-agent-local-deployment/</guid>
      <pubDate>Wed, 26 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 IBM 官方发布、模型卡与代码，分析 Granite 4.2 的 3B/8B/30B、推理档位、128K/512K 上下文、工具调用、许可和企业本地部署边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>ArmorOCR 对抗性文字识别：文档数字员工为何需要区域证据与双路复核</title>
      <link>https://clawsphere.io/research/armorocr-adversarial-document-agent-security/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/armorocr-adversarial-document-agent-security/</guid>
      <pubDate>Tue, 25 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 ArmorOCR 官方权重、代码和论文，分析对抗性视觉文字如何绕过 OCR，拆解区域证据、双路识别、策略护栏与人工复核的企业文档安全架构。</description>
      <category>治理合规</category>
    </item>
    <item>
      <title>新款 Mac Studio 发布：M5 Ultra 512GB 能否成为企业本地 AI 工作站</title>
      <link>https://clawsphere.io/research/mac-studio-m5-ultra-local-ai-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/mac-studio-m5-ultra-local-ai-2026/</guid>
      <pubDate>Tue, 25 Aug 2026 00:00:00 GMT</pubDate>
      <description>分析 M5 Max 与 M5 Ultra Mac Studio 的统一内存、带宽、Thunderbolt 5 集群和本地大模型能力，以及企业采购前必须验证的边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>SingGuard 多模态安全护栏：企业数字员工如何把动态策略变成可审计控制</title>
      <link>https://clawsphere.io/research/singguard-multimodal-agent-guardrail-enterprise-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/singguard-multimodal-agent-guardrail-enterprise-governance/</guid>
      <pubDate>Sun, 23 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于百灵 SingGuard 官方权重、仓库与技术报告，分析运行时策略、多模态审核、快慢推理和本地部署边界，给出数字员工输入、输出与工具调用的治理架构。</description>
      <category>治理合规</category>
    </item>
    <item>
      <title>OpenAI ZDR 与 Private Safety Processing：企业 Agent 的数据治理架构</title>
      <link>https://clawsphere.io/research/openai-zdr-private-safety-processing-enterprise-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/openai-zdr-private-safety-processing-enterprise-governance/</guid>
      <pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 OpenAI 8 月 19 日官方预告与 API 数据控制文档，拆解零数据保留、跨会话安全信号、端点例外和企业数字员工的状态、审计与采购边界。</description>
      <category>治理合规</category>
    </item>
    <item>
      <title>数字员工开始自主付款后：AgentCore Payments 的企业架构与治理清单</title>
      <link>https://clawsphere.io/research/agentcore-payments-enterprise-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/agentcore-payments-enterprise-governance/</guid>
      <pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 AWS AgentCore Payments 8 月 18 日 GA，拆解机器支付、钱包委托、会话预算、商户校验、幂等补偿与审计边界，给出企业上线验收框架。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>实时数字员工如何断线续跑：Google ADK Live 会话治理与幂等设计</title>
      <link>https://clawsphere.io/research/google-adk-live-session-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/google-adk-live-session-governance/</guid>
      <pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 Google ADK 8 月 17 日更新，拆解实时 Agent 的会话恢复、音频结束、后台工具收尾与至少一次执行风险，给出企业上线架构和验收清单。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>Kimi K3 企业部署研究：2.8T 开放权重如何改变模型路由、成本与许可决策</title>
      <link>https://clawsphere.io/research/kimi-k3-enterprise-deployment-routing/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/kimi-k3-enterprise-deployment-routing/</guid>
      <pubDate>Sat, 15 Aug 2026 00:00:00 GMT</pubDate>
      <description>Kimi K3 提供 2.8T 开放权重、104B 激活参数和 1M 上下文，但开放权重不等于低成本私有化。本文拆解企业 Agent 的集群成本、模型路由、许可与采购边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>DeepSeek V4 Pro 正式版：企业 Agent 如何评估推理档位、峰谷价格与 API 兼容</title>
      <link>https://clawsphere.io/research/deepseek-v4-pro-enterprise-agent-routing/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/deepseek-v4-pro-enterprise-agent-routing/</guid>
      <pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate>
      <description>DeepSeek V4 Pro 0813 正式版带来可调推理档位、1M 上下文与峰谷计费；Responses API 已确认支持，但仍有无状态与静默忽略字段等兼容边界。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>2026 端侧模型大全与本地部署指南：48 个模型家族怎么选</title>
      <link>https://clawsphere.io/research/edge-model-local-deployment-guide-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/edge-model-local-deployment-guide-2026/</guid>
      <pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate>
      <description>系统比较 Qwen、Gemma、MiniCPM、LFM、GLM、ArmorOCR、InternVL、Llama、Mistral 等端侧模型的设备层级、工作负载、内存误区和企业选型方法。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>营销数字员工如何落地：从 AI 洞察到持续执行闭环</title>
      <link>https://clawsphere.io/research/enterprise-marketing-digital-employee-closed-loop/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/enterprise-marketing-digital-employee-closed-loop/</guid>
      <pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate>
      <description>结合 OpenAI 的 Zapier 案例与 Google Ads、Analytics 新能力，拆解企业营销数字员工从洞察、决策、执行到复盘的闭环架构与治理边界。</description>
      <category>企业实践</category>
    </item>
    <item>
      <title>企业 Agent 如何防提示注入？OpenAI 与 Anthropic 给出的共同警告</title>
      <link>https://clawsphere.io/research/agent-prompt-injection-security/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/agent-prompt-injection-security/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>从网页、邮件、文档和 MCP 工具返回值解释提示注入与数据外泄风险，并给出内容隔离、最小权限、审批和审计的纵深防御方案。</description>
      <category>治理合规</category>
    </item>
    <item>
      <title>Anthropic Agent 研究趋势 2026：自治、专业知识与可信治理</title>
      <link>https://clawsphere.io/research/anthropic-agent-research-trends-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/anthropic-agent-research-trends-2026/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>基于 Anthropic 对 Claude Code 真实使用数据的研究，分析 Agent 自治时长、领域专家价值、监督方式、提示注入和企业治理趋势。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>Coding Agent 真能提升企业生产力吗？解读 OpenAI 与 Anthropic 的最新证据</title>
      <link>https://clawsphere.io/research/coding-agents-enterprise-productivity-evidence/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/coding-agents-enterprise-productivity-evidence/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>综合 OpenAI Codex 使用研究与 Anthropic Claude Code 会话研究，解释哪些生产力信号可信、哪些不能外推，以及企业如何设计自己的对照实验。</description>
      <category>企业实践</category>
    </item>
    <item>
      <title>企业 Agent Runtime 选型：SDK、框架、运行时与 Agent Builder 怎么分</title>
      <link>https://clawsphere.io/research/enterprise-agent-runtime-selection/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/enterprise-agent-runtime-selection/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>从 Agent SDK、编排框架、托管 Runtime 和低代码 Agent Builder 四层拆解企业技术栈，并给出权限、状态、观测、评测与迁移清单。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>NVIDIA、AMD 与国产 AI 算卡：企业采购到底该比较什么</title>
      <link>https://clawsphere.io/research/enterprise-ai-accelerator-selection/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/enterprise-ai-accelerator-selection/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>从芯片、软件栈、模型适配、集群、功耗、供应与迁移成本比较 NVIDIA、AMD、昇腾、寒武纪、昆仑芯、壁仞和摩尔线程。</description>
      <category>企业实践</category>
    </item>
    <item>
      <title>企业智能办公 Agent 全景：M365、WPS、Dumate、钉钉与飞书如何选</title>
      <link>https://clawsphere.io/research/enterprise-ai-workplace-landscape/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/enterprise-ai-workplace-landscape/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>比较 Microsoft 365 Copilot、WPS Comate、百度办公搭子 Dumate、钉钉 AI 助理与飞书 Aily，解释办公入口、知识、权限和数字员工的选型逻辑。</description>
      <category>行业趋势</category>
    </item>
    <item>
      <title>2026 基座模型生态地图：14 个主流家族与企业多模型路由</title>
      <link>https://clawsphere.io/research/foundation-model-ecosystem-map-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/foundation-model-ecosystem-map-2026/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>核验 GPT、Claude、Gemini、Qwen、DeepSeek、Llama、ERNIE、混元、豆包、Kimi、MiniMax、GLM、Mistral 与 Grok 的当前主线。</description>
      <category>模型观察</category>
    </item>
    <item>
      <title>Agent 越来越自治，人应该怎么管？从逐步审批到持续监督</title>
      <link>https://clawsphere.io/research/human-oversight-for-autonomous-agents/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/human-oversight-for-autonomous-agents/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>结合 Anthropic 的真实自治研究与 OpenAI 的 Codex 控制实践，给出企业设计可见、可干预、风险分级和可追责监督机制的方法。</description>
      <category>企业实践</category>
    </item>
    <item>
      <title>长时 Agent 怎么评测？从聊天正确率升级到环境任务成功率</title>
      <link>https://clawsphere.io/research/long-running-agent-evaluation/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/long-running-agent-evaluation/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>解释为什么长时 Agent 不能只用问答基准评测，并给出环境、工具、轨迹、人工干预、外部状态和成本组成的企业评测方法。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>MCP 进入企业之后怎么管？连接器、身份、权限与供应链清单</title>
      <link>https://clawsphere.io/research/mcp-enterprise-governance/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/mcp-enterprise-governance/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>MCP 降低了 Agent 连接工具的成本，也扩大了权限和供应链风险。本文给出企业 MCP 注册、身份、审批、版本和审计治理框架。</description>
      <category>治理合规</category>
    </item>
    <item>
      <title>OpenAI Agent 研究趋势 2026：从 Codex 到长时数字工作</title>
      <link>https://clawsphere.io/research/openai-agent-research-trends-2026/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/openai-agent-research-trends-2026/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>梳理 OpenAI 2026 年 Agent 研究主线：长时任务、并行 Agent、技能与自动化、环境评测、安全边界，以及从编程向企业知识工作的扩展。</description>
      <category>Agent 研究</category>
    </item>
    <item>
      <title>OpenAI 与 Anthropic 的 Agent 路线有什么不同？企业应该看懂的六个信号</title>
      <link>https://clawsphere.io/research/openai-vs-anthropic-enterprise-agent-strategy/</link>
      <guid isPermaLink="true">https://clawsphere.io/research/openai-vs-anthropic-enterprise-agent-strategy/</guid>
      <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
      <description>从产品载体、长时任务、真实使用研究、安全治理、企业控制和开放协议六个维度比较 OpenAI 与 Anthropic 的 Agent 战略。</description>
      <category>行业趋势</category>
    </item>
  </channel>
</rss>