Kimi K3’s release last week was the point where “the best open-weight model” and “the best model, period” stopped being far apart. The catch, as the XDA piece put it, is that you still cannot run K3 at home. What you can run is the previous tier: Qwen, DeepSeek, Llama, Gemma, and Mistral all have open weights, and the Android tooling around them has caught up in the last twelve months.
We tested seven Android apps that let you chat with open-source models, either on-device or against a server you control. Some are polished consumer apps that happen to expose an open-weights option. Others are unapologetic tinkerers’ tools that let you side-load a GGUF and see what a phone SoC can actually do.
What to look for in an open AI chat app
- On-device vs remote. On-device means the model runs on your phone’s NPU or CPU. Remote means you send messages to a server (yours or the vendor’s) that runs the model.
- Weights source. Consumer apps hide it. Tinkerer apps let you download or side-load specific GGUF or MLC files.
- Context window. Smaller phone-runnable models often cap at 8K or 32K tokens. Server-backed apps run the full advertised context.
- Whether the app requires an account. DeepSeek and Perplexity do. MLC Chat and PocketPal do not.
- Privacy path. On-device inference never leaves the phone. Everything else needs a policy check.
Quick comparison
| App | Best for | On-device? | Free plan | Standout |
|---|---|---|---|---|
| DeepSeek | DeepSeek-V3 chat | No | Free with account | Free access to a top-tier model |
| Google Gemini | Gemini Nano on-device | Yes (Nano) | Free | Runs on Tensor G3+ chips |
| Google AICore | System-level Gemini Nano | Yes | Free, built-in | Powers other apps’ on-device AI |
| MLC Chat | On-device Llama/Mistral/Gemma | Yes | Free, open-source | Universal GGUF/MLC runner |
| PocketPal | Local model tinkering | Yes | Free, open-source | Runs Hugging Face GGUF files |
| ChatterUI | Character chat on any backend | Yes and remote | Free, open-source | Works with local or hosted APIs |
| Perplexity | Open-weight answers with sources | No | Free tier plus Pro | Sonar model over open weights |
The seven apps
1. DeepSeek, best free access to a frontier open model
DeepSeek’s Android app is a clean chat interface for DeepSeek-V3, one of the strongest open-weight models available today. Chat, DeepThink for extended reasoning, and web search are the three modes. The app itself is free and the underlying model is served without a paid subscription for personal use.
Where it falls short: the model runs on DeepSeek’s infrastructure, not on your phone. You need an account. Chats are used to improve the service unless you opt out in settings.
Pricing: Free with account.
Download: Aptoide · Google Play
Bottom line: The pick when you want frontier open-model quality without hosting anything.
2. Google Gemini, best on-device chat with a mainstream polish
Google Gemini’s Android app ships with Gemini Nano running fully on-device on Tensor G3 and later chips. On other phones the app falls back to cloud Gemini. Nano is limited to short-context tasks like summarisation and quick replies, but the app is the smoothest mainstream way to see on-device inference in action.
Where it falls short: full-featured chat still round-trips to the cloud. Nano’s on-device capabilities are quiet in the UI, so it is not always obvious what is local and what is remote.
Pricing: Free. Google One AI Premium subscription $19.99 per month adds Gemini Advanced and larger context.
Download: Aptoide · Google Play
Bottom line: The polished pick if you already have a Pixel 8 or newer.
3. Google AICore, best system-level on-device backbone
AICore is not a chat app on its own. It is the system service that hosts Gemini Nano and exposes it to other apps that want on-device generation. Once installed, other tools can call into it without loading their own model weights. Recorder summaries and TalkBack image descriptions already rely on it.
Where it falls short: no user interface. It is a background service, and its usefulness is entirely a function of the apps that call into it.
Pricing: Free, bundled with Pixel and Samsung Galaxy AI.
Download: Aptoide · Google Play
Bottom line: Keep it up to date if you use any Google on-device AI features.
4. MLC Chat, best universal on-device runner
MLC LLM is the community project that turned llama.cpp-style local inference into a proper Android app. MLC Chat ships as an open-source APK that runs Llama, Mistral, Phi, Gemma, and Qwen variants converted into MLC format. The catalog inside the app lets you download a model over Wi-Fi and chat with it entirely offline.
Where it falls short: models take gigabytes of storage. Inference on mid-range phones is slow enough that longer answers feel like typing on a 90s terminal. No cloud fallback.
Pricing: Free and open-source.
Download: GitHub release APK. Not currently on Aptoide or Google Play.
Bottom line: The go-to when you want to prove open-weight chat can run without a server.
5. PocketPal, best for casual model tinkering
PocketPal is a friendly open-source Android app for testing local models. It downloads GGUF weights from Hugging Face and runs them via a llama.cpp core, with a simple chat UI and per-conversation settings for temperature, top-p, and context window. It is the least intimidating way to run a small local model.
Where it falls short: performance is a function of the model. A 3B Q4 quant is usable, an 8B is slow, anything bigger will not fit in RAM. There is no cloud fallback.
Pricing: Free and open-source.
Download: Aptoide · Google Play
Bottom line: The right pick to poke at Qwen 3B or Gemma 2B without setting up a server.
6. ChatterUI, best for connecting any backend
ChatterUI is an Android chat interface that talks to whatever backend you point it at: a local model via llama.cpp, a self-hosted Ollama instance on your LAN, or a hosted API. It is a bring-your-own-model app that keeps the interface consistent while you swap providers underneath.
Where it falls short: setup requires knowing where your model actually lives. Not a fit if you want the app to pick a model for you.
Pricing: Free and open-source.
Download: GitHub releases APK. Also available on some third-party APK stores.
Bottom line: The pick when you already have an Ollama, LM Studio, or hosted endpoint and want a mobile front end.
7. Perplexity, best open-weight answer engine with sources
Perplexity uses a mix of proprietary and open-weight models under the hood, including its Sonar family which is trained on Llama. The Android app pairs those models with live web search and returns answers with sources cited inline. It is not on-device, but it is one of the strongest consumer-facing uses of open-weight foundations.
Where it falls short: model choice is limited on the free tier. Sonar Pro and DeepSeek modes need a Perplexity Pro subscription.
Pricing: Free with account. Perplexity Pro around $20 per month unlocks model selection.
Download: Aptoide
Bottom line: The pick for open-weight assistants that also do research with sources.
How to pick the right one
- Just want a strong chat with an open model: DeepSeek.
- Have a Pixel 8+ and want on-device polish: Google Gemini.
- Want to prove local inference works on a phone: MLC Chat or PocketPal.
- Already run Ollama at home: ChatterUI.
- Want cited answers over open weights: Perplexity.
FAQ
Can I run Kimi K3 on my phone? Not usefully. K3 is a large mixture-of-experts model that needs server-class hardware. Smaller Qwen and Llama variants are what actually fit on phones today.
What is the best free open-source model I can chat with on Android? DeepSeek is the strongest free chat experience. For on-device, Gemma 2B or Qwen 2.5 3B in MLC Chat or PocketPal are the practical picks.
Do local models drain the battery quickly? Yes. Sustained inference pulls hard on the SoC. Expect 5 to 10 percent battery for a long chat with a 3B model.
Is DeepSeek safe to use in Europe? DeepSeek has been the subject of privacy debates in several EU countries. Its Android app processes chats on servers in China. Choose accordingly.
What is Google AICore actually for? It is a system service that hosts Gemini Nano so other apps can call it without shipping their own weights. Install it if you want on-device AI features in other Google apps to work.
