Unlocking Remote AI: Securely Access Self-Hosted DeepSeek

Why Self-Hosting AI is Great—But Accessing It Isn’t
Self-hosting large language models (LLMs) like DeepSeek is gaining traction for good reasons: privacy, cost efficiency, and performance. But there’s one big hurdle—how do you access your AI remotely without exposing it to the open internet?
Let’s say you’ve got DeepSeek running locally on your workstation. Now you want to query it from your laptop while traveling, connect it to an edge device, or share it securely with a team. Traditional networking options—VPNs, port forwarding, or cloud relays—are either too complex, too slow, or just plain insecure.
Enter RingLink.
RingLink is a decentralized, fully meshed networking solution that allows you to securely connect devices without a centralized server. Unlike traditional P2P tools, RingLink offers direct peer connections when possible but also provides fallback encrypted relays for cases where direct communication is blocked. In this guide, we’ll show you how to:
✅ Set up DeepSeek locally
✅ Create a private RingLink network
✅ Expose DeepSeek over RingLink for remote access
✅ Access your AI from anywhere
DeepSeek R1 Model Hardware Requirements
Model Size | Minimum CPU | Min RAM (CPU‑only) | Recommended GPU (VRAM) | Disk Storage (~GB) | Inference Speed – tokens/s (CPU; RTX 4090; A100) |
---|---|---|---|---|---|
1.5B | Modern 4‑core CPU (2013+) | 8 GB | RTX 3060 (12 GB, uses ~0.7 GB VRAM) | ≈1.1 | 5 t/s (CPU); **100** t/s (4090); ~80 t/s (A100) |
7B | Modern multi-core CPU (4–6 cores) | 16 GB | RTX 3070 (8 GB, uses ~3.3 GB VRAM) | ≈4.7 | ~2 t/s (CPU); ~50 t/s (4090); ~40 t/s (A100) |
14B | High-performance CPU (6+ cores) | 32 GB | RTX 3080 (10 GB+, uses ~6.5 GB VRAM) | ≈9.0 | ~1–2 t/s (CPU); ~60 t/s (4090); ~50 t/s (A100) |
32B | High-performance CPU (8+ cores) | 64 GB | RTX 4090 (24 GB, uses ~14.9 GB VRAM) | ≈20 | <1 t/s (CPU); ~30 t/s (4090); ~25 t/s (A100) |
70B | Workstation-grade CPU (12+ cores) | ~128 GB | 2× RTX 4090 (24 GB ea., ~32.7 GB total) | ≈43 | ≪1 t/s (CPU); ~2 t/s (1×4090 with CPU offload); ~10 t/s (A100 80GB) |
671B | Multi-socket server CPUs (e.g. dual EPYC) | ≈384–512 GB | 16× A100 80GB (~1342 GB total VRAM) | ≈400 | ~5 t/s (CPU only); n/a on single 4090/A100 |
- 1.5B model runs on basic hardware – no dedicated GPU or VRAM required.
- 7B model benefits from ≥8 GB VRAM but can run on CPU (slowly).
- 14B+ models require high-performance CPUs and large RAM capacity.
- 70B model can run with 128 GB RAM and CPU offloading but is best on multi-GPU setups.
- 671B model cannot run on a single GPU – requires multi-GPU clusters with at least 1.3 TB VRAM.
Step-by-Step Guide
Step 1: Install and Run DeepSeek Locally

1. Install Ollama for Easy Model Management
curl -fsSL https://ollama.com/install.sh | sh
This starts a local DeepSeek server at localhost:11434
.
🔍 Troubleshooting: If another service is using 11434
, specify a different port:
env OLLAMA_HOST=127.0.0.1:11435 ollama serve
2. Download the DeepSeek Model
ollama pull deepseek-r1:1.5b
Step 2: Install and Set Up RingLink
1. Install RingLink & Join the Network
- For Linux:
- Visit RingLink Dashboard and click “Add Linux Server” to generate an installation script.
- Run the generated command:
bash <(curl -s https://assets.ring.link/install.sh) --config-key rl-config-key-{config-key} --user-id {user-id} && ringlink-cli connect {network-id}
- For macOS, Windows, and other platforms:
- Download the appropriate installer from the official RingLink website: RingLink Downloads

2. Verify Connection Status
After installation, check if RingLink is running and connected to the network:
ringlink-cli status
If successfully connected, it should display “Running” and indicate the associated network.
3. Join the Network from Other Devices
Once joined, devices in the network can communicate securely without port forwarding.

🔍 If you experience connection issues:
- Run
ringlink-cli status
to check if the network is active.
- Restart RingLink with
systemctl restart ringlink
.
- Ensure your firewall isn’t blocking the connection.
Step 3: Expose DeepSeek Over RingLink
After setting up your DeepSeek instance, you can expose it over RingLink’s Edge Tunnel (ELB), making it securely accessible without opening public ports.
1. Verify Your DeepSeek Instance
Before exposing your model, ensure it is running correctly:
curl http://localhost:11434/api/generate -d '{
"model": "deepseek-r1:1.5b",
"prompt":"Why is the sky blue?"
}'
2. Exposing DeepSeek via RingLink Edge Tunnel (ELB)
Step 1: Create an ELB Instance
1. Visit RingLink ELB Management.
2. Click “Create”, then confirm the ELB configuration.
Step 2: Configure ELB Rules for DeepSeek
1. Find the newly created ELB instance and click “Details”.

2. Set the Source to your DeepSeek server:
- Source: Select Device, then choose your DeepSeek instance (e.g., Mac.lan (264263db0d)).
- Port: Enter 11434 (or your actual DeepSeek port).

3. Assign an external port for public access.
Example:
- External Port: 21434
- Your Auto-Generated ELB Endpoint: motionless-advertising.go-elb.com:21434
📌 This is your unique external access point for DeepSeek. Use it to securely connect without exposing your local network.

4. Choose TCP as the protocol and save the configuration.
3. Access DeepSeek Over the Internet
Now that DeepSeek is exposed through RingLink ELB, you can access it securely from any device:
curl http://motionless-advertising.go-elb.com:21434/api/generate -d '{
"model": "deepseek-r1:1.5b",
"prompt":"Why is the sky blue?"
}'
📢 Make sure to replace the “motionless-advertising.go-elb.com:21434
” with your real ELB endpoint.
💡 Performance Considerations:
- RingLink uses a fully meshed architecture, allowing direct connections whenever possible.
- If a direct P2P connection is blocked, RingLink’s Secure Orb Relay ensures stable connectivity with minimal latency.
- Compared to VPNs, RingLink avoids centralized coordination servers, reducing network overhead.
Use Cases: Where This Setup Works Best
- Secure AI Access from Anywhere – Query your local AI model remotely without security risks.
- Team Collaboration on Self-Hosted AI – Give teammates controlled access to your DeepSeek instance.
- Distributed AI Processing – Link multiple self-hosted AI instances for more computing power.
- Edge AI Networking – Deploy self-hosted AI models across multiple devices without relying on cloud services.
With this optimized approach, securely accessing your self-hosted DeepSeek AI from anywhere is simple, efficient, and safe.