DeepFlow launches a trusted AI mobile architecture, opening a new paradigm for Web3 intelligent assistants

In 2026, AI is influencing and taking over various industries, demonstrating the ability to redefine everything. Following the viral success of ByteDance’s Doubao Phone, innovations in the AI assistant field continue to emerge: Zhipu open-sourced the AutoGLM framework, and the rise of AI assistant applications like Cowork has also attracted widespread market attention. These consecutive events reveal the potential for AI to reshape the entire mobile ecosystem. As the first AI-native smartphone, Doubao Phone achieves screen perception, user intent interpretation, and cross-application task execution through operating system-level proxies, gaining popularity among users and sparking extensive discussions about ecological disruption and privacy security. Meanwhile, AutoGLM, as an open-source framework, further democratizes AI capabilities, turning autonomous task execution from concept to reality. AI + smartphones are unstoppable in redefining the interaction model between humans and intelligent agents, rewriting existing market rules.

As a pioneer in the integration of AI and Web3, DeepFlow has launched a new trusted AI smartphone architecture that automates complex Web3 interactions while achieving ultimate privacy protection through Trusted Execution Environment (TEE) and cloud-edge hybrid architecture. By seamlessly connecting with exchanges, dApps, and underlying protocols, DeepFlow effectively addresses long-standing issues of data privacy and operational efficiency, and accelerates the industry’s evolution toward agentic, privacy-first decentralized systems. On the eve of Artificial General Intelligence (AGI), the strategic value of this architecture is especially prominent: building trustworthy computing-based intelligent agents to help users achieve seamless autonomous operations within the Web3 ecosystem while avoiding privacy leaks.

Technical Foundation: Deeply adapting GUI intelligent agents to Web3 scenarios

DeepFlow’s core architecture leverages multimodal reasoning and graphical interface navigation capabilities, reconstructing leading GUI proxy models into a unified Web3 proxy layer that autonomously executes tasks such as browser dApp interactions, wallet operations, and smart contract calls. It adopts an innovative “cloud-edge hybrid intelligent architecture” to balance performance, privacy, and cost. This design not only enhances operational efficiency but also strategically positions fully automated AI assistants on the eve of AGI, ensuring user intent is securely and autonomously executed in complex Web3 environments.

Edge Side: Lightweight Models and Privacy Sentinels

Deploy highly optimized lightweight AI models on user terminals (browser plugins, mobile apps, dedicated hardware modules). Their core responsibilities include:

• Privacy data preprocessing and recognition: locally detecting and desensitizing all sensitive inputs (such as private key fragments, transaction passwords, identity information, asset data), ensuring raw privacy data never leaves the user device.
• Initial intent understanding and task decomposition: parsing user natural language commands into standardized, structured sub-tasks.
• Integration with Trusted Execution Environment (TEE): performing critical sensitive operations like signatures and key management within isolated TEE secure enclaves (Intel SGX and ARM TrustZone), providing hardware-level security guarantees.

This edge-side design ensures low-latency responses and minimizes privacy risks, providing reliable support for local autonomy of fully automated AI assistants.

Cloud Side: CPU TEE + GPU NVTrust Trusted Computing Solution and Decentralized Computing Network

In the cloud, DeepFlow connects high-performance large models (such as Auto-GLM, Mai-UI) as the “brain,” responsible for complex strategy analysis, market trend interpretation, cross-protocol routing optimization, and other tasks requiring high computing power and broad knowledge. More importantly, the cloud introduces advanced privacy-preserving interaction mechanisms: receiving task data that has been desensitized and encrypted at the edge. Using secure multi-party computation (MPC), homomorphic encryption (HE), and other cryptographic techniques, computations are performed without decrypting the original data, returning execution strategies or results.

DeepFlow’s cloud architecture emphasizes the hardware foundation of trusted computing, combining CPU TEE and GPU NVTrust technologies to form a comprehensive confidential computing framework. This not only breaks through the limitations of edge models but also establishes a resilient, privacy-first execution pipeline, laying a trust foundation for agent-driven Web3 interactions.

CPU TEE Integration: Utilizing Intel SGX and similar CPU TEE technologies, cloud processors execute sensitive computations in isolated environments. TEE creates secure enclaves, ensuring code and data are protected from external access during runtime, including from cloud administrators or potential attackers. This is especially critical when handling Web3 protocol routing and strategy optimization, such as in multi-party collaboration scenarios, where TEE allows joint computations on encrypted data, preventing privacy leaks.

GPU NVTrust Integration: To meet the high computational demands of large AI models, DeepFlow adopts NVIDIA’s NVTrust framework, supporting GPU-level confidential computing. NVTrust extends traditional TEE into the GPU domain, leveraging architectures like NVIDIA Hopper, Blackwell, or Vera Rubin to create protected zones within GPU memory. Sensitive AI inferences (such as multimodal task decomposition or market prediction) can be executed within these zones without exposing model weights or user data. NVTrust ensures verifiability of the computation process through hardware root trust and remote attestation, even in shared cloud environments, preventing side-channel attacks.

The advantages of this combined approach include:

• Performance and security balance: CPU TEE handles control flow and logic tasks, while GPU NVTrust accelerates parallel, compute-intensive operations like neural network inference, achieving millisecond response times.
• End-to-end privacy chain: from desensitized data at the edge to cloud-side processing, combined with MPC and HE, ensuring privacy at the level of Zero-Knowledge Proofs.
• Compliance and scalability: conforming to GDPR and Web3 privacy standards, supporting future integration of more NVIDIA hardware such as NVSwitch for large-scale distributed computing.

The following table compares key differences between traditional cloud architectures and DeepFlow’s cloud trusted computing solution:

This trusted computing solution not only enhances the reliability and efficiency of the cloud side but also provides a solid trust foundation for the Web3 ecosystem, avoiding the common assumption of “trusting the cloud provider” found in traditional environments. This is crucial for strategic deployment on the eve of AGI, helping fully automated AI assistants achieve efficient, privacy-secure cloud expansion.

Innovative Proxy Model: AI as the Middle Layer of Web3

DeepFlow redefines AI agents as decentralized proxies, abstracting the complexity of Web3 protocols and providing natural language or API-driven intent execution interfaces. The proxies built by DeepFlow possess high autonomy, capable of aligning intents through multimodal inputs (screenshots, transaction logs) and executing closed-loop operations. Transitioning from passive tools to active intelligent agents, DeepFlow injects an efficient execution middle layer into the Web3 ecosystem, accelerating the transition to “Agentic Web3.” The strategic value of this model lies in bridging user intent with complex Web3 operations, ensuring a central role in the future era of smart technology.

Agent Autonomy: Combining Chain-of-Thought (CoT) reasoning with MCP (Model-Callable Protocol) tool invocation, proxies can flexibly switch between GUI simulation and native Web3 APIs (js, Solana RPC, etc.).

Privacy-First Design: Proxies default to differential privacy techniques and TEE secure enclaves to anonymize user data, preventing information leaks during cross-protocol interactions. In the cloud, NVTrust further enhances this design, ensuring GPU-accelerated tasks (such as real-time market analysis) remain private.

Connecting Exchanges and Protocols: Proxies act as universal connectors, enabling automated liquidity operations between CEX and DEX, reducing operational friction in CDeFi hybrid modes, supporting real-time portfolio rebalancing, and strictly isolating KYC data.

API-First Product Approach: Initially launched as lightweight API suites for quick B-side partner integration. Future plans include expanding to SDKs for deeper platform integration and leveraging cloud trusted computing solutions to enhance security.

Strategic Deployment on the Eve of AGI

DeepFlow will further support intent-driven architectures and emerging AI-native blockchain paradigms, collaborating with top exchanges, leading infrastructure providers, and dApp developers for joint R&D and strategic partnerships. Through the trusted computing solutions of cloud CPU TEE + GPU NVTrust, it not only addresses current privacy pain points but also paves the way for the future Web3 Agentic era, ensuring user privacy while enjoying AI convenience. On the eve of AGI, DeepFlow embeds fully automated trusted AI assistants as core engines of Web3, reshaping traffic logic and leading the industry toward smarter and safer evolution.