AI News AI资讯 2h ago Updated 1h ago 更新于 1小时前 49

Perplexity Ships Portable Computer on NVIDIA DGX Spark: Local Harness, OS-Enforced Sandbox, and Zero Per-Token Cost for Local Steps Perplexity在NVIDIA DGX Spark上推出便携式计算机:本地编排、OS强制沙箱与本地步骤零Token成本

Perplexity released Portable Computer, a local-first agentic AI platform that runs the full agent harness, orchestrator, planner, tool router, and post-trained models directly on NVIDIA DGX Spark hardware The system uses an escalation gate design where every task starts on-device, and cloud model usage (15+ options) requires explicit per-step user approval after PII classification Portable Computer ships as an integrated system (not just a local chat app), supporting Qwen 3.8 27B and PPLX 27B mo Perplexity发布Portable Computer,将完整agentic AI平台(harness、orchestrator、planner、tool router)本地运行于NVIDIA DGX Spark,本地任务零per-token费用 采用"本地优先+混合云"架构:所有任务始于本地设备,需云推理时经PII检查后逐步骤请求用户批准,仅发送单个步骤至15+云模型 支持Qwen 3.8 27B和PPLX 27B模型,配备OS级强制沙箱隔离,代码/工具执行受限进程、文件系统和网络访问 基准测试:53任务Local Knowledge Work Bench达85.4%(PPLX 27B)

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Perplexity released Portable Computer, a local-first agentic AI platform that runs the full agent harness, orchestrator, planner, tool router, and post-trained models directly on NVIDIA DGX Spark hardware
  • The system uses an escalation gate design where every task starts on-device, and cloud model usage (15+ options) requires explicit per-step user approval after PII classification
  • Portable Computer ships as an integrated system (not just a local chat app), supporting Qwen 3.8 27B and PPLX 27B models with OS-enforced sandboxing for code and tool execution
  • Benchmarks show PPLX 27B achieving 85.4% on the 53-task Local Knowledge Work Bench, significantly outperforming Pi (77.6%) and Hermes (74.0%) on identical hardware
  • The hybrid local-cloud approach on Terminal Bench 2.1 achieved 73.0% at ~$0.415/rollout versus 82.4% at ~$0.65 for Claude Opus 5 alone, narrowing the frontier gap at lower cost

Why It Matters

Portable Computer represents a significant step toward enterprise-deployable local AI agents that balance data residency requirements with cloud-powered reasoning capabilities. The explicit escalation gate with PII checking addresses a critical concern for regulated industries (finance, healthcare, government, defense) where cloud inference is blocked by contractual or compliance constraints. For AI practitioners, this demonstrates a practical architecture for local-first agentic systems that can economically handle repo-scale migrations and long verification loops without per-token charges.

Technical Details

  • Architecture: Full agent stack (harness, orchestrator, planner, tool router, post-trained models) runs locally on NVIDIA DGX Spark with GB10 superchip, 128 GB RAM, and 1 TB storage minimum; alternative support for RTX GPUs with 24+ GB VRAM on DGX OS or Ubuntu
  • Model Options: Qwen 3.8 27B (3-bit quantized, 17.4 GB download, 32 GB RAM required) and PPLX 27B (Perplexity's post-trained variant tuned for its harness); NVIDIA Nemotron 3.5 Lightning (30B MoE, 4-bit, 19 GB, 36 GB RAM) coming soon; bring-your-own model support included
  • Sandboxing & Tool Execution: OS-enforced sandbox restricts processes, filesystem paths, and network access; tool execution is disabled rather than silently downgraded if sandbox unavailable; Gmail, Outlook, Slack, and GitHub connectors route through local orchestrator as compact CLI tools
  • Context Management: Engineered around 27B model context limits (Qwen 3.8 advertises 260K tokens but degrades past ~100K) by keeping system prompt/toolset small, loading specialized skills on demand, and compacting stale context mid-run
  • Escalation Protocol: Before any cloud call, harness selects relevant context, runs PII classifier, displays exactly what would leave the machine to user; approved step routes to one of 15+ cloud models; remote adviser returns text guidance only with no direct access to local files, tools, or conversation

Industry Insight

  • Enterprise AI procurement will increasingly factor in hardware gate requirements: Portable Computer's hard hardware gate (GB10-class or RTX 24GB+ VRAM) creates a clear segmentation where only enterprises with existing NVIDIA workstations or well-funded startups can deploy, reinforcing the "machine is the price of entry" dynamic for local AI
  • Hybrid local-cloud architectures with explicit escalation gates will become a standard pattern for regulated industries: The PII-check-then-approve workflow demonstrates how to balance cost efficiency (local processing at zero marginal cost) with capability (cloud frontier models for web/reasoning tasks), a pattern likely to be adopted across finance, legal, healthcare, and government AI deployments
  • Local agent benchmarks are emerging as a meaningful evaluation paradigm: Perplexity's Local Knowledge Work Bench (53 tasks spanning research, financial analysis, document creation) and results showing 85.4% with PPLX 27B provide a template for evaluating local-first agentic systems beyond traditional chat benchmarks, with implications for how AI practitioners assess deployment readiness

TL;DR

  • Perplexity发布Portable Computer,将完整agentic AI平台(harness、orchestrator、planner、tool router)本地运行于NVIDIA DGX Spark,本地任务零per-token费用
  • 采用"本地优先+混合云"架构:所有任务始于本地设备,需云推理时经PII检查后逐步骤请求用户批准,仅发送单个步骤至15+云模型
  • 支持Qwen 3.8 27B和PPLX 27B模型,配备OS级强制沙箱隔离,代码/工具执行受限进程、文件系统和网络访问
  • 基准测试:53任务Local Knowledge Work Bench达85.4%(PPLX 27B),Terminal Bench 2.1混合模式73.0%成本仅$0.415/rollout,远低于Claude Opus 5的$0.65
  • 硬件门槛高:需GB10超级芯片+128GB内存+1TB存储,或RTX 24GB+ VRAM,主要面向企业/中型团队及资金充裕的AI原生初创公司

为什么值得看

本文展示了本地优先agentic AI的工程化落地路径,为数据敏感行业提供了兼顾成本、性能与合规的可行方案。其混合推理架构和PII分级审批机制,为AI安全部署提供了可复用的设计范式。

技术解析

  • 一体化本地部署架构:Portable Computer将本地模型、推理引擎、agent harness、工具沙箱和应用连接器打包为单一系统,消除传统本地AI部署中手动搭建推理服务器和连接工具的复杂性,用户只需选择Qwen 3.8 27B或PPLX 27B即可开箱即用。
  • 混合推理与PII分级审批:系统采用"本地处理为主、云推理为辅"策略,当步骤需要实时网络或前沿推理时,orchestrator暂停并展示即将离机的数据(经PII分类器处理),用户逐步骤批准后仅发送该步骤至15+云模型,远程顾问仅返回文本指导且不接触本地文件、工具或对话。
  • 小模型上下文优化工程:针对Qwen 3.8 27B标称260K-token窗口但实际100K后性能衰减的问题,系统保持精简系统提示和工具集、按需加载专业技能、将连接器暴露为紧凑CLI工具而非完整MCP定义,并在运行中压缩过时上下文。
  • OS级强制沙箱:代码和工具调用在操作系统层面强制隔离,限制进程、文件系统路径和网络访问;若沙箱不可用则直接禁用工具执行而非静默降级,确保安全边界不被绕过。
  • 硬件与量化规格:DGX Spark需GB10超级芯片+128GB内存+1TB存储;Qwen 3.8 27B以3-bit量化(17.4GB下载,32GB RAM);Nemotron 3.5 Lightning 4-bit量化(19GB,36GB RAM);其他系统需DGX OS或Ubuntu(ARM/x64)+ RTX 24GB+ VRAM。

行业启示

  • 本地优先agentic AI成为企业级部署新范式:金融、法律、医疗、政府和国防等数据敏感行业可借此实现合规的AI自动化,本地处理消除数据外泄风险,混合云架构保留前沿能力。
  • 成本-性能平衡的混合推理模式具有商业价值:Terminal Bench 2.1显示混合模式(73.0%)以64%成本($0.415 vs $0.65)逼近Claude Opus 5(82.4%),为长验证循环和仓库级迁移提供经济可行的替代方案。
  • 小模型工程优化正在释放本地AI实用潜力:通过上下文压缩、按需技能加载和紧凑工具定义,27B级模型在本地可达85.4%基准准确率,证明"够用就好"的本地模型+智能编排架构可减少对超大模型的依赖。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent GPU GPU Inference 推理 Deployment 部署 Product Launch 产品发布