Phos vs Google AI Edge Gallery, Maid, and Off Grid

How to read this comparison

This guide compares product direction and documented capability, not audited binary equivalence or a universal privacy score. Primary project repositories and official documentation were reviewed on August 1, 2026. Features can change after publication. Test the installed version, model, phone, and route before trusting sensitive data.

The central distinction is purpose. A model gallery, direct model client, offline multimodal suite, and personal assistant can all run local AI while optimizing different experiences.

Comparison table

Cells summarize publicly documented project direction, not a guarantee about every release or phone.

Phos: assistant continuity

Phos starts from repeated personal work: think through a decision, draft a difficult message, study, plan a day, remember selected context, and organize projects. The model is one layer under that experience. Local mode aims to keep sensitive prompts on Android; local-server mode uses hardware the user runs; BYOK uses the user's provider relationship. Route status should remain visible.

The tradeoff is deliberate scope. Phos V1 is Android-first and does not promise every offline modality, automatic cloud sync, autonomous agents, company-hosted inference, or a universal model laboratory. Users who primarily want to test experimental models or generate offline media may prefer a more specialized reference app.

Google AI Edge Gallery: official optimized showcase

Google's repository describes AI Edge Gallery as a showcase for on-device ML and generative-AI use cases and current platform support. It is a strong place to try supported models and understand the LiteRT ecosystem. Google uses the app in its developer material for optimized local models, and current releases evolve quickly.

The question for a personal-assistant buyer is whether the current release supplies the desired persistence, memory, projects, privacy controls, and everyday workflows. A gallery can demonstrate excellent inference without trying to become a long-term companion. That difference is about purpose, not quality.

Maid: model and endpoint flexibility

Maid's repository documents local llama.cpp GGUF files plus remote connections to providers and Ollama. It appeals to users who already understand model files, system prompts, endpoints, and provider credentials. Bring-your-own-model flexibility can expose new model families quickly when the runtime supports them.

That freedom also gives the user more responsibility for model source, license, quantization, template, RAM, and provider policy. Compare current release provenance, credential handling, chat storage, and route labels. A broad integration list does not mean every route has the same privacy boundary.

Off Grid: offline modality breadth

Off Grid's public project describes an offline suite spanning text generation, vision, transcription, image generation, tool calling, and document analysis. For a user who wants several local media capabilities on a powerful phone, that product direction can be more important than assistant memory or BYOK flexibility.

Each modality introduces model downloads, storage, native libraries, permissions, and device constraints. Verify which feature and model support the Android device, what optional downloads require, and how generated or imported files are retained. "All local" still needs ordinary mobile storage and export controls.

Privacy comparison without shortcuts

Run the same verification routine for each candidate. Start core use without an account if that matters. Download one model, force-stop, enter airplane mode, and generate. Inspect permissions, data-safety disclosures, memory, deletion, exports, notifications, and backups. If the app offers online providers or servers, identify the destination before sending.

Open source helps reviewers trace these paths, but connect repository tags and signing to the installed release. Do not assume a familiar publisher or many GitHub stars proves the exact privacy property you need. Evidence should be version- and route-specific.

Which should you choose?

Choose Phos when the priority is an approachable personal assistant with no Phos login, explicit inference routes, local memory control, and practical workflows. Choose AI Edge Gallery when the priority is Google's supported on-device model experience and experimentation. Choose Maid when direct GGUF and endpoint choice matter most. Choose Off Grid when offline multimodal breadth is the decisive feature.

Install two rather than relying on a generic ranking. Run the same ten prompts, note model load, answer utility, thermals, context, storage, deletion, and failure behavior. Keep the one that solves the repeated job while making its boundaries easiest to verify.