Most AI apps lock you into one model per conversation. If you want to compare how Llama handles a question versus Qwen, you open two apps or two browser tabs, paste the same prompt, and compare side by side. If you want to start with a fast model and switch to a smarter one for the hard parts, you lose your context and start over.
That is not how you would use AI if there were no artificial barriers. You would use the right model for each question, in the same conversation, without thinking about it.
Off Grid lets you do exactly that. Switch between any model - on your phone or on your network - at any point in a conversation. The chat history stays. The context carries over. You just change which brain is answering.
How it works
Off Grid gives you access to models from two sources:
On your phone. Smaller models that run directly on your hardware. Qwen 3.5 0.8B, 2B, Phi-4 Mini, SmolLM3. These load into your phone's memory and run inference on the CPU/GPU. No network needed.
On your network. Larger models running on Ollama or LM Studio on your Mac or PC. Off Grid auto-discovers these servers and shows you every available model. Qwen 3.5 9B, Llama 3.1, Mistral, anything your server has loaded.
All of these models show up in one model selector. Tap to switch. Continue chatting. That is it.
Off Grid showing both on-device and network models in a single app - iOS, Android, Ollama, and LM Studio all working together.
Why you would want to switch models mid-chat
Speed versus depth. Start with the 2B model for brainstorming. It responds in milliseconds on-device. When you have an idea worth developing, switch to the 9B on your Mac for the detailed work. You get the speed of a small model for exploration and the depth of a large model for execution, without splitting your thought process across two apps.
Different strengths. Models are not all good at the same things. A code-specialized model handles programming questions better. A general model handles writing better. A multilingual model handles translation better. In one conversation, you might need all three. Instead of three apps, you have one conversation with three models taking turns.
Testing your prompts. If you write prompts for work or for a product, you need to know how different models respond. Paste a prompt, try it with Qwen 3.5 9B, switch to Llama 3.1, switch to Mistral. Same conversation, immediate comparison.
Privacy escalation. You are asking general questions and the on-device model is fine. Then the conversation gets personal - medical, legal, financial. Switch to the on-device model for those parts. Nothing leaves your phone. Then switch back to the remote model when you are past the sensitive section.
Graceful degradation. You are at home using the 9B on your Mac. You leave the house. The remote model becomes unavailable. Instead of losing your conversation, switch to the on-device 2B and keep going. The quality drops, but the conversation does not stop.
What models work
Any GGUF model on-device, and any model exposed through an OpenAI-compatible API remotely. In practice, this means:
On-device (from Off Grid's model browser, filtered by your phone's RAM):
- Qwen 3.5 0.8B, 2B (best all-rounders for mobile)
- Phi-4 Mini (strong reasoning for its size)
- SmolLM3, SmolVLM (vision capable)
- Any .gguf file you import manually
Remote (from Ollama or LM Studio on your network):
- Qwen 3.5 4B, 9B (the sweet spot for consumer hardware)
- Llama 3.1 8B, 70B (if your machine can handle it)
- Mistral, DeepSeek, Gemma, anything your server has
- Code-specific models like CodeLlama or DeepSeek Coder
Off Grid does not care where the model comes from or what format it is in. If it can talk to it, you can use it.
This costs nothing
Every model mentioned above is free to download. Ollama is free. LM Studio is free. Off Grid is free and open source. Your phone and your Mac are already paid for. Your WiFi is already on.
Qwen 3.5 9B, released March 2026, outperforms OpenAI's GPT-OSS-120B on reasoning and language benchmarks at 13 times smaller. It runs comfortably on a MacBook with 16GB of RAM. You have a multi-model AI setup that rivals cloud subscriptions costing $240+ per year, running on hardware you already own, and it costs nothing.
Beyond chatting
Multi-model access is not just for conversations. Off Grid also supports:
Projects with RAG. Attach documents to a project and query them with any model. Use a fast model for quick lookups, switch to a larger model for synthesis across multiple documents.
Tool calling. Models that support function calling can use built-in tools - web search, calculator, date/time, device info. The larger remote models are significantly better at deciding when and how to use tools, so having both small and large models available in the same app means you always have the option.
Vision. Attach an image or point your camera. Models with vision support analyze what they see. SmolVLM and Qwen3-VL run on-device. Larger vision models can run on your server.
Voice. On-device Whisper transcription. Dictate, transcribe locally, send to whatever model you want.
Where Off Grid is heading
The long-term vision is a personal AI operating system where you do not have to think about models at all. The system knows your devices, knows what compute is available, and routes each question to the right model automatically. Small factual question? On-device, instant. Complex analysis? Routed to your desktop GPU. You just talk to your AI and the infrastructure handles itself.
That is where we are heading. Today, the model switching is manual. Tomorrow, it is automatic. We are building it in the open.
Join the Off Grid Slack from our GitHub.
Try it
- GitHub (1,000+ stars, 10,000+ downloads in 4 weeks)
- Android: GitHub Releases
- iOS: App Store (search "Off Grid AI")
Download the app, set up a couple of on-device models, connect to your Ollama or LM Studio server, and you have more AI models at your fingertips than any single subscription service offers. All free. All private. All on your own hardware.
Off Grid is built by the team at Wednesday Solutions, a product engineering company with a 4.8/5.0 rating on Clutch across 23 reviews.
Top comments (0)