Figure 计划本周发布 Hark AI 助手

Figure plans to release Hark this week as a proactive AI assistant. It is taking shape as a personal AI system built around memory, proactive work, and computer use. What we have spotted in development suggests a personalized home screen where users can create widgets connected to email, calendars, and other services. The same surface would bring together recurring tasks, inbox panels, weather, and daily briefs designed to decide whether something needs attention or Hark should stay quiet.

This direction closely matches what Hark has already announced publicly. The company calls Hark “personal intelligence” rather than a traditional chatbot, combining persistent memory with the ability to act across the web. Its Terms of Service confirm that Hark can remember context across conversations, access connected email, calendars, documents, and messaging services, operate a browser, make transactions when authorized, and create documents and web pages.

Hark has already demonstrated part of this stack through Handoff, its computer-use agent announced in August. On X, Hark showed Handoff ordering gifts, booking appointments, and completing end-to-end web tasks, positioning it primarily around everyday consumer work rather than coding. Hark said Handoff topped several browser-use benchmarks and would form part of its broader personal assistant product.

The unreleased product appears to extend Handoff much further. Memory could cover people, places, projects and voice conversations, while users may be able to inspect or delete stored information. A Markdown-based “soul” document could control personality, while reusable browser skills would let users define procedures for specific websites. The text also references research, spreadsheets, documents, image generation, website building, MCP connectors, phone access, and messaging integrations.

The captured model configuration points to a mix of specialized models. A model labeled Claude Opus 5.5 is assigned to the main assistant, task agents, and website building, while Gemini 3.1 Flash Image is configured for image generation. Browser-login settings reference Qwen3.6-27B and Gemini 3.5 Flash Lite, and browser challenge handling names GPT-6 Astra. These are internal configuration identifiers; they do not independently establish publicly available models or confirm which model handled an actual request.

The company raised more than $700 million at a $6 billion valuation, is developing its own foundation models and hardware, partnered with AT&T for standalone connected AI devices, and secured gigawatt-scale NVIDIA Vera Rubin capacity.

The App Store previously listed Hark Pro for September 30, while Hark’s website still says the product is launching soon. The emerging Pro, Pro², Pro³, and Pro⁴ plans, with references to 5× and 20× usage options, suggest higher tiers may mainly increase usage limits, with automatic token top-ups also in the works.

Sources and related context

• Hark Terms of Service: Supports: Hark's terms describe persistent memory, connected services, browser actions, authorized transactions, content creation, and optional automatic usage top-ups.
• Introducing Hark Handoff: Supports: Hark's August announcement describes Handoff's end-to-end browser tasks, its reported benchmark results, and its role in the broader personal intelligence platform.
• Hark announces its NVIDIA partnership: Supports: The announcement confirms gigawatt-scale NVIDIA Vera Rubin capacity supporting Hark's personalized AI platform, foundation models, and bespoke hardware.