By Sagar Shankaran, Founder of CallSphere
Samsung's Galaxy S26 runs a full agentic AI system locally on the Exynos 2600 chip, handling complex multi-step tasks offline with no cloud dependency.
Key takeaways
Samsung's Galaxy S26, unveiled at Mobile World Congress Barcelona on March 3, 2026, represents a fundamental shift in how AI agents operate on consumer devices. For the first time, a mainstream smartphone ships with a fully autonomous agentic AI system that runs entirely on-device — no internet connection required for complex multi-step task execution.
The system, branded "Galaxy AI Agent," is powered by Samsung's custom Exynos 2600 system-on-chip, which features a dedicated Neural Processing Unit (NPU) capable of 45 TOPS (trillion operations per second) and 12GB of on-device model memory. This hardware foundation enables a 7-billion-parameter language model to run inference at approximately 30 tokens per second — fast enough for real-time conversational interaction and tool use.
"We drew a hard line in the design process: the agent must be fully functional in airplane mode," said TM Roh, President of Samsung's Mobile Experience division, during the launch keynote. "If your AI assistant stops working the moment you lose signal, it is not really an assistant. It is a proxy."
Samsung's implementation differs architecturally from cloud-based agent systems in several critical ways.
flowchart LR
INPUT(["User intent"])
PARSE["Parse plus<br/>classify"]
PLAN["Plan and tool<br/>selection"]
AGENT["Agent loop<br/>LLM plus tools"]
GUARD{"Guardrails<br/>and policy"}
EXEC["Execute and<br/>verify result"]
OBS[("Trace and metrics")]
OUT(["Outcome plus<br/>next action"])
INPUT --> PARSE --> PLAN --> AGENT --> GUARD
GUARD -->|Pass| EXEC --> OUT
GUARD -->|Fail| AGENT
AGENT --> OBS
style AGENT fill:#4f46e5,stroke:#4338ca,color:#fff
style GUARD fill:#f59e0b,stroke:#d97706,color:#1f2937
style OBS fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b
style OUT fill:#059669,stroke:#047857,color:#fff
Quantized Foundation Model: The on-device model is a heavily quantized (4-bit GPTQ) version of Samsung's proprietary Gauss 3 language model. While the full Gauss 3 model runs at 70 billion parameters in data centers, the on-device variant has been distilled to 7 billion parameters with task-specific fine-tuning that preserves performance on mobile-relevant tasks while dramatically reducing compute requirements.
Local Tool Registry: The agent has access to a pre-registered set of device-level tools — calendar management, email composition, settings modification, app navigation, file management, camera control, and communication APIs. Each tool is exposed to the model through structured function definitions stored on-device, with no need for API calls to external servers.
On-Device RAG: Samsung has implemented a local retrieval-augmented generation system that indexes user data (contacts, messages, emails, documents, photos with OCR text) into an on-device vector store. This enables the agent to answer questions about personal data without any information leaving the device.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
Persistent Memory Store: The agent maintains a local SQLite-backed memory system that records user preferences, task history, and learned patterns. Over time, the agent adapts its behavior — learning, for example, that a user always wants meeting reminders 15 minutes early, or that they prefer certain apps for specific tasks.
Samsung published benchmark comparisons against cloud-based alternatives that have drawn both praise and skepticism from the tech community.
On Samsung's internal "AgentBench-Mobile" benchmark, which measures success rate on 500 multi-step mobile tasks, the Galaxy S26's on-device agent scored 73.2% — compared to 81.4% for Google's cloud-based Gemini agent on Pixel 10 and 79.8% for Apple Intelligence on iPhone 17 Pro. The gap narrows considerably for tasks that do not require web search or access to external APIs.
For latency, the on-device system has a clear advantage. First-token latency averages 180 milliseconds, compared to 400-800 milliseconds for cloud-based alternatives depending on network conditions. For multi-step tasks requiring several agent reasoning loops, cumulative latency savings are substantial.
"The latency advantage is real and matters more than people think," said Andrej Karpathy in a post on X. "Agent systems run multiple inference passes per task. Shaving 300ms off each pass means a 10-step task completes 3 seconds faster. That is the difference between feeling instant and feeling sluggish."
Samsung is marketing the on-device approach heavily on privacy grounds, positioning it against competitors whose AI features rely on cloud processing.
All agent interactions, memory stores, and personal data indexing happen exclusively on the device. Samsung has published a technical whitepaper detailing the security architecture, including hardware-level encryption for the agent's memory store and a secure enclave for processing sensitive data like financial information and health records.
This approach addresses a growing consumer concern. A February 2026 survey by Pew Research found that 67% of US adults were "somewhat or very concerned" about AI assistants sending personal data to cloud servers. Among respondents under 35, 58% said they would pay a premium for AI features that work entirely on-device.
The European Data Protection Board (EDPB) has also signaled support for on-device AI processing, with a January 2026 opinion paper suggesting that on-device AI systems may face lighter regulatory requirements under the AI Act compared to cloud-based alternatives.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
Samsung released the Galaxy AI Agent SDK alongside the device, enabling third-party developers to register their apps as tools that the on-device agent can invoke. Early partners include Spotify (music control and playlist creation), Uber (ride booking), and several banking apps in South Korea and Europe.
The SDK uses a structured schema format similar to OpenAI's function calling specification, making it relatively straightforward for developers already building for cloud-based agent systems to port their tool definitions to Samsung's on-device framework.
"We had our Uber integration working in about two days," said a developer from Uber's mobile team speaking at the MWC developer session. "The schema format is nearly identical to what we already use for Google's agent integration. The main difference is that everything happens locally — the agent calls our app's local API rather than a cloud endpoint."
Samsung's move puts pressure on Apple and Google, both of which have relied primarily on cloud-based approaches for their most capable AI features.
Apple Intelligence, introduced with iOS 18, processes some tasks on-device but routes complex multi-step operations to Apple's Private Cloud Compute infrastructure. Google's Gemini agent on Pixel devices is almost entirely cloud-dependent, though Google has hinted at on-device capabilities coming with the Pixel 11 later in 2026.
Qualcomm, whose Snapdragon 8 Elite 2 chip powers many competing Android flagships, announced at MWC that it is working with Meta to bring a similar on-device agent capability using Llama 4 Scout to devices powered by its platform, expected in Q3 2026.
Initial reviews from tech publications have been largely positive about the on-device approach, while noting limitations. The agent occasionally struggles with complex reasoning chains that require more than 6-7 sequential steps. Tasks involving web interaction obviously require connectivity, though the agent can queue actions and execute them once a connection is restored.
Battery impact is the most cited concern. Running the on-device model at full capacity draws approximately 4 watts from the NPU, which Samsung estimates reduces battery life by roughly 15% during heavy agent usage. The system includes intelligent scheduling that routes less time-sensitive agent tasks to periods when the device is charging.
Pre-orders for the Galaxy S26 have reportedly exceeded those of the S25 by 40% in South Korea, with Samsung attributing much of the demand to the on-device AI capabilities.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
On-device voice LLMs are now real. What Apple Intelligence, Gemini Nano, and Phi-4 ship in 2026 — and what they cannot do yet.
Run a 3-node k3s cluster at the network edge to slash voice-agent first-token latency below 250ms. ServerlessLB, MetalLB, NodeLocalDNS, and tuned WebRTC ports.
Set up CDN and edge caching for your AI agent's static assets, API responses, and pre-computed results to reduce global latency with proper cache headers, edge functions, and geographic optimization.
Edge AI runs inference directly on devices, eliminating cloud latency and enabling real-time decisions. Learn how on-device AI works and where it delivers the most value.
Implement privacy-preserving techniques in AI agent systems including differential privacy for data aggregation, federated learning for distributed model training, on-device processing, and compliance with GDPR and CCPA requirements.
Master TensorFlow Lite for deploying AI agent models on Android and iOS devices, including model conversion, quantization strategies, and real-world integration patterns.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco