Thursday, 24 Sep 2026
Subscribe to AIWatcher
AIWatcher
  • Home
  • News

    OpenAI slots two cheaper GPT-6 models beneath Astra

    By
    AIWadmin

    Researcher hijacks Meta’s Muse agent through a hidden setting

    By
    AIWadmin

    ChatGPT learns to take voice orders for office chores

    By
    AIWadmin

    YouTube hands creators an AI agent for titles and thumbnails

    By
    AIWadmin

    Listeners can now rewrite the Spotify algorithm in plain words

    By
    AIWadmin

    Anthropic’s Claude agents flag a new enzyme family in viral DNA

    By
    AIWadmin
  • Articles

    Data shop Snorkel AI banks $350M as labs stockpile training sets

    By
    AIWadmin

    Ema banks $77M to push AI employees into the back office

    By
    AIWadmin

    US military command randomises routes to outfox enemy models

    By
    AIWadmin

    Kyutai teaches a speech model to do arithmetic out loud

    By
    AIWadmin

    Nokia hands developers a no-training route to calibrated answers

    By
    AIWadmin

    An open model sorts eight overlapping speakers in real time

    By
    AIWadmin
  • Spotlight

    DeepMind keeps cloud memory encrypted behind device-held keys

    By
    AIWadmin

    Anthropic trims Opus costs and speeds output in a mid-cycle refresh

    By
    AIWadmin

    Cisco Talos builds a fingerprint library for AI-driven malware

    By
    AIWadmin

    Google parks idle agents in a new open source runtime

    By
    AIWadmin

    OpenAI gives mathematicians a voice but hands them no brake

    By
    AIWadmin

    NVIDIA teaches its robotics stack to take instructions from agents

    By
    AIWadmin
  • About
    • Mission
    • Services
    • Contact
  • Newsletter
  • Shop
    • All Items
    • By Category
      • Hats
      • T-Shirts
    • Cart
  • 🔥
  • Alignment
  • Classification
  • Distillation
  • Explainability
  • Hallucination
  • Legal/Compliance
  • Medical
  • NLM
  • Mobility
  • Research
  • Robotics
  • Safety
  • Startups
  • Prompt
  • Python
  • RAG
  • RLHF
  • Token
  • Vision
Font ResizerAa
AIWatcherAIWatcher
  • Home
  • News
  • Articles
  • Spotlight
  • About
  • Newsletter
  • Shop
Search
  • Home
  • News
  • Articles
  • Spotlight
  • About
    • Mission
    • Services
    • Contact
  • Newsletter
  • Shop
    • All Items
    • By Category
    • Cart
Have an existing account? Sign In
Follow US
© 2022 Foxiz News Network. Ruby Design Company. All Rights Reserved.
News

Perplexity packs its agent computer into a DGX Spark

A local-first build of the agentic platform runs models on-device with zero per-token cost.

AIWadmin
Last updated: August 26, 2026 8:32 pm
AIWadmin
ByAIWadmin
Global AI news & information.
Follow:
Share
SHARE

Perplexity has taken its agentic Computer platform fully local with a release called Portable Computer. Everything ships as one system: the agent harness, the orchestrator, the planner, the tool router, and the post-trained models all run on the device. The target box is an NVIDIA DGX Spark, and local steps carry no per-token charge.

Two models ship at launch: Qwen 3.8 27B and PPLX 27B, Perplexity’s own post-trained variant, with NVIDIA Nemotron 3.5 Lightning expected later. The whole setup sits behind an OS-level sandbox. Processes, filesystem paths, and network access are all confined to it. If the sandbox cannot start, tool execution is disabled rather than downgraded.

The setup is local-first but not local-only. When a step requires live web data or frontier reasoning, the orchestrator pauses and asks for approval, running a PII classifier over the context and showing users exactly what would leave the machine. Escalated steps go to one of 15+ cloud models, which receive only text guidance and no direct access to local files or tools.

Benchmarks on Perplexity’s own 53-task work bench put the Qwen-powered build at 82.6% and the PPLX variant at 85.4%. Terminal Bench 2.1 results tell a hybrid story: a fully local run scored 59.6% at near-zero marginal cost, escalation lifted that to 73% at about $0.415 per rollout, and Claude Opus 5 alone hit 82.4% at roughly $0.65.

Linux users on Pro, Max, and Enterprise plans get access first, with Windows support in September. The hardware bar is a GB10-class machine or an RTX GPU with 24GB of VRAM.

TAGGED:AI AgentsDGX SparkEdge ComputingLocal AINvidiaPerplexity
SOURCES:MarkTechPost
Share This Article
Email Copy Link Print
ByAIWadmin
Follow:
Global AI news & information.
Previous Article Meta’s MetaRoCE lets AI Ethernet survive packet loss
Next Article McKinsey finds AI profits still flat as spending climbs

You Might Also Like

News

Google AI answers gain ground as publishers lose traffic

By
AIWadmin
News

Robot data startup Mecka draws Sequoia at a $500M value

By
AIWadmin
News

Disney creates a CTO role and hands it to Character.AI’s chief

By
AIWadmin
News

Harvey previews its first in-house legal model for long case runs

By
AIWadmin
AIWatcher
Facebook Twitter Youtube Linkedin Rss

Global AI News and Information
AIWatcher is your definitive source for AI updates worldwide, from Silicon Valley to Shanghai.
Our industry coverage keeps you in the loop with the latest news and trends shaping the future of AI.

Quick Links
  • News
  • Articles
  • Spotlight
  • Events
About Us
  • Mission
  • Services
  • Contact
  • Privacy Policy
  • Legal
© 2026 AIWatcher. All Rights Reserved.