Google's Gemma 4 Replaces Claude Pro in Homelab Setup, Ending $20/Month Subscription

At a glance:

• Google's Gemma 4 e4b model replaces Claude Pro in local AI workflows
• Tailscale enables mobile access to local AI via Open WebUI
• Annual savings of $240 achieved through self-hosted deployment

The Shift from Cloud to Local AI

Shekhar Vaidya, a tech journalist and former XDA contributor, recounts his decision to abandon Claude Pro after discovering Google's Gemma 4. For months, Vaidya paid $20 monthly for Claude Pro while his RTX 4070 Ti GPU remained largely idle. His AI usage primarily involved casual tasks on his phone—summarizing content, brainstorming, and quick Q&A—rather than complex coding or research. Local models like 7B or 24B failed to match cloud performance due to small context windows and slow inference. Gemma 4's e4b variant, however, offered a 128K context window, native vision support, and MoE architecture that fit his 12GB VRAM GPU. While it didn't surpass Claude Pro in reasoning, it eliminated the need for a subscription by running entirely locally.

The shift wasn't about outperforming Claude Pro but about practicality. Vaidya's mobile use cases didn't require frontier models. Gemma 4's free, private deployment on his homelab provided equivalent functionality without rate limits or privacy concerns. The setup involved Ollama for model management, Open WebUI for the interface, and Tailscale for secure cross-device access. This combination replaced his $240 annual Claude Pro subscription with a self-hosted solution that leveraged existing hardware.

Technical Implementation

The deployment process was simpler than expected. Vaidya used Ollama's Windows installer to run Gemma 4:e4b, a 9.6GB model optimized for his GPU. The setup required only two commands: pulling the model via ollama pull gemma4:e4b and launching Open WebUI with Docker. The latter involved mapping ports and setting environment variables to enable local access. Initial issues with Windows Subsystem for Linux (WSL) memory allocation were resolved by adjusting the .wslconfig file. Once stable, Open WebUI mirrored Claude's interface, supporting chat history, image uploads, and model selection.

Gemma 4's MoE architecture was key to efficiency. Only active model segments loaded during inference, keeping GPU utilization at 57% during typical tasks. This contrasted with Claude Pro's cloud dependency, which incurred costs and latency. Vaidya noted that while Gemma 4 lacked Claude's advanced reasoning, it sufficed for 90% of his mobile workflows. The model's 128K context window also handled longer documents—a limitation in smaller local models.

Mobile Integration via Tailscale

Tailscale was critical to making local AI accessible on his phone. By routing Open WebUI through Tailscale's private network, Vaidya accessed his homelab from anywhere without public exposure. The setup required Tailscale on both his PC and iPhone, but once configured, the experience felt seamless. He added Open WebUI as a Progressive Web App (PWA) to his home screen, eliminating the need to type IP addresses. This integration allowed him to send prompts via mobile data or Wi-Fi, with responses generated on his GPU at home.

The security benefits were notable. All traffic remained within Tailscale's encrypted network, bypassing CGNAT restrictions. Vaidya emphasized that local AI on mobile wasn't just about cost savings but about control. Unlike cloud services, he could modify the setup, ensure data privacy, and avoid dependency on third-party APIs. Tailscale's role in bridging his homelab and mobile device was pivotal, transforming a desktop tool into a portable solution.

Why This Matters

Vaidya's experience reflects a broader trend: local AI is becoming viable for everyday use. While frontier models like Claude Pro or Opus 4.7 still excel in complex tasks, self-hosted solutions address privacy, cost, and accessibility concerns. Gemma 4's success on consumer hardware demonstrates that powerful models don't require enterprise infrastructure. This shift could democratize AI access, allowing users to leverage high-performance models without subscription fees.

However, limitations remain. Gemma 4's MoE architecture and 9.6GB size restrict it to mid-range GPUs. Larger models or tasks requiring advanced reasoning may still favor cloud solutions. Vaidya himself acknowledges that local AI hasn't replaced cloud models entirely but has carved a niche for routine use. The rise of tools like Ollama and Tailscale lowers the barrier to entry, making self-hosting feasible for non-experts.

Future Outlook

The integration of local AI with mobile devices via Tailscale could expand further. As models like Gemma 4 evolve, their context windows and capabilities may close the gap with cloud alternatives. Vaidya suggests watching for improvements in MoE efficiency and smaller model variants. Additionally, cross-platform tools that simplify homelab setups could attract more users. For now, his setup serves as a proof of concept—showing that practical, private AI is achievable without sacrificing performance.

Conclusion

By combining Gemma 4, Ollama, and Tailscale, Vaidya eliminated a recurring expense while enhancing privacy and hardware utilization. His journey underscores the growing potential of local AI, particularly for users prioritizing cost control and data sovereignty. While not a universal replacement for cloud services, self-hosted solutions are gaining traction as models become more efficient and tools more user-friendly.