Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM/LM Playground Review: Open-Source Offline LLM Chat for Android
Mobile & Edge LLMs

LM Playground Review: Open-Source Offline LLM Chat for Android

·8 min read·By Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI
Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

LM Playground (version 1.10.1 as of 16 October 2026, last updated on Google Play on 8 October 2026) needs no account and is built on llama.cpp, the open-source engine that executes the models on the phone. Models are downloaded once from Hugging Face and then chat, document questions, photo input, and dictation work offline. Only two things go online by design: model downloads and the optional web search and page-fetch tools, which are off by default.

LM Playground is a free, MIT-licensed Android app by the independent developer Andriy Druk that runs open-weight language models in GGUF format directly on the phone through llama.cpp. Beyond chat it offers on-device document Q&A, photo input for vision models, reasoning-model display, optional tools, and voice dictation. This review is based on the Google Play listing, the official website, the privacy policy, and the public GitHub repository, checked on 16 October 2026; PromptQuorum has not tested the app hands-on.

LM Playground Review: Open-Source Offline LLM Chat for Android

Key Takeaways

  • Developer: Andriy Druk, an independent Android developer; the project has been public on GitHub since March 2024 and is a one-person project as far as the sources show.
  • Price and license: free with no account; the source code is under the MIT license.
  • Models: a built-in catalog (the website says 40+) from Qwen, Gemma, Llama, GPT-OSS, Nemotron, Phi-4, Granite, Mistral, DeepSeek, and Liquid AI, plus your own GGUF files.
  • Platform: Android only; the sources mention no iPhone, desktop, or web version.
  • Recency: Google Play shows the last update on 8 October 2026 and GitHub shows release 1.10.1 on 12 October 2026, so both dates are recent.

📍 In One Sentence

LM Playground turns an Android phone into a private place to chat with open-weight models, question your own documents, and show photos to vision-capable models.

💬 In Plain Terms

You install it from Google Play, download a model once, and then chat without an account or an internet connection; the web tools and model downloads are the only parts that need the network.

📌Note: This review is based on the Google Play listing, the official website, the privacy policy, and the public GitHub repository (README, license, releases, and release notes), checked on 16 October 2026. PromptQuorum has not tested or benchmarked the app.

What Is LM Playground?

LM Playground is an Android chat app that runs language models locally instead of calling a cloud service. According to its README, you download models, load them in one tap, and chat offline; the inference engine is llama.cpp and the models are GGUF files from Hugging Face. Models are not bundled, so the first run needs a network connection.

The developer, Andriy Druk, is listed on Google Play under that name and publishes the project in his own GitHub account. The project is unrelated to products with similar names, such as the Playground tools of cloud AI vendors.

Get It

LM Playground is a store app: install it from Google Play, the one channel the project and its website point to. GitHub is for source code and release notes; the releases page carries no APK file.

Platform
Where to get it
AndroidGoogle Play (free; minimum Android version not shown)
Websitelmplayground.app
Privacy policyPrivacy policy for LM Playground
Source codeGitHub (MIT)

Version and recency as verified on 16 October 2026: the latest GitHub release is 1.10.1, published on 12 October 2026 (the website header still shows 1.10.0). Google Play lists the app as updated on 8 October 2026 but does not show its build number, so the Play build may be older than the newest GitHub release.

How to Get Started

Setup follows the order described in the README and Play listing; PromptQuorum has not run these steps.

  1. 1
    Install the app
    Why it matters: Install LM Playground from Google Play; no account or API key is required according to the listing.
  2. 2
    Download a model on Wi-Fi
    Why it matters: Pick a model from the built-in list or load your own GGUF file. The smallest model is 267 MB, models of 4B to 8B parameters benefit from 8 GB of RAM or more per the Play listing, and downloads resume after interruptions.
  3. 3
    Chat, attach a file, or send a photo
    Why it matters: Start a chat, attach a document and ask about it, or send an image to a vision-capable model. Document search and dictation each download a small helper model once on first use.
  4. 4
    Optionally turn on tools
    Why it matters: Tools are off by default and can be enabled per model in Settings. Leave them off to keep every chat on the device.

Which Models Does It Offer?

Qwen 3.5 / Qwen 3

Sizes:
0.8B, 2B, 4B / 0.6B, 1.7B, 4B
Notes:
Qwen 3.5 entries accept images

Gemma 4 / Gemma 3n / Gemma 3

Sizes:
E2B, E4B, 12B / E2B, E4B / 1B, 4B
Notes:
Vision on Gemma 4 E4B and Gemma 3 4B per the README

GPT-OSS

Sizes:
20B
Notes:
Needs a high-end phone with plenty of memory per the Play listing

Llama 3.2 / 3.1

Sizes:
1B, 3B / 8B
Notes:
From Meta

Granite, Nemotron, Phi-4, DeepSeek R1 Distill

Sizes:
Granite 4.x 3B to 8B; Nemotron 4B; Phi-4 mini 3.8B; R1 Distill 1.5B, 7B
Notes:
DeepSeek R1 and Nemotron show their thinking steps

LFM2.5, MiniCPM5, SmolLM3, Spark-X2.5, Ministral 3

Sizes:
LFM2.5 350M to 2.6B; LFM2.5 VL 450M to 3B; Ministral 3 3B, 8B
Notes:
Ministral 3 and LFM2.5 VL accept images

Entries come from the repository README on 16 October 2026; most use Q4_K_M quantization, Qwen 3.5 uses Q3_K_M, and GPT-OSS ships in its native MXFP4 format. The Play listing names a shorter list of ten families, and the catalog changes with each release.

Features Confirmed by the Sources

Every item below comes from the README, the website, the release notes, or the Play listing; none has been independently tested.

  • Document Q&A. Attach a PDF, Word, EPUB, HTML, Markdown, or plain-text file; on-device semantic search finds the relevant passages and hands them to the model. It downloads a 278 MB embedding model once.
  • Vision input. Attach a gallery photo or take one with the camera and ask a vision-capable model to describe it or read its text.
  • Voice dictation. Added in version 1.10.0: tap the microphone and speak; recognition runs on the phone in 35 languages after a one-time 718 MB speech model download. It appears on the website and in the release notes but not in the Play description.
  • Reasoning display. Thinking steps of models such as DeepSeek R1 and Nemotron appear in a collapsible section with an adjustable thinking budget.
  • Control and history. Per-model settings (context size, temperature, Top-P, Top-K, Min-P, repetition penalty, seed), reusable system prompts, and a searchable chat history with pin, rename, and delete.
  • Tools. Capable models can search the web, fetch a page, and run JavaScript in a sandbox; each tool is off by default.
  • Background generation and storage. Replies keep running when you leave the app, models can live in any folder or on an SD card, and the interface is translated into 28 languages.

Privacy and Network Use

Core use is offline by design: the website claims zero network calls at inference, and the Google Play Data safety section declares "No data collected" and "No data shared with third parties." The privacy policy says the developer collects no personal or non-personal information.

That policy took effect on 16 March 2024 and describes Internet access only as the permission that lets the app download models. The web search and page-fetch tools came later; the sources do not say which search service receives a query, so treat those two tools as the part that goes online and keep them off for confidential chats.

📌Note: The source code is public, so these claims can be checked by anyone who builds the app, but PromptQuorum has not audited the code or the network traffic.

Trade-Offs: Benefits vs. Limitations

Free and open source

What it means in real use:
No account, no subscription, and the MIT-licensed code can be read and built.
Limitation / caveat:
A single developer maintains it; the sources state no support commitment.

Bring your own GGUF

What it means in real use:
Load model files from any folder next to the built-in catalog.
Limitation / caveat:
Play reviews we read call loading external models hit-and-miss; these are individual user reports.

Documents, photos, voice

What it means in real use:
One app covers file questions, image input, and dictation, all on the device.
Limitation / caveat:
Quality depends on the small model you load, and helper models add downloads.

Active development

What it means in real use:
Frequent releases add new models and fixes, such as the 1.10.1 vision-crash fix.
Limitation / caveat:
Fast change means the Play build, the website, and GitHub can show different versions.

Optional web tools

What it means in real use:
Models can look up current pages when you enable the tools.
Limitation / caveat:
Goes online, and a Play reviewer reported the model looping while "thinking" with web search on.

Who Should Use It

  • Android users who want a free, auditable offline chat app. The MIT license and public repository make it checkable.
  • People who want to question their own files privately. Document Q&A with on-device search is a headline feature.
  • Tinkerers who bring their own models. Custom GGUF loading and per-model sampling controls are built for that.

What We Could Not Verify

  • Hands-on performance. PromptQuorum did not run the app, so speed, battery use, and answer quality are not assessed.
  • Which build Google Play serves. The listing shows an update date but no version number or minimum Android version.
  • Which search service the web tool uses. The sources read do not name it.
  • Whether the privacy policy covers the web tools. The policy text predates them and was not found updated.
  • Not for iPhone, Mac, or Windows users. The sources describe an Android-only app.

Competitors and Alternatives

PocketPal AI

Platforms:
iOS, Android
Price / license:
Free / MIT
Key difference:
Open-source on-device chat client with its own model library
Articles about PocketPal AI (10)

+24 more not shown

Off Grid

Platforms:
iOS, Android, Mac, Windows
Price / license:
Freemium / MIT
Key difference:
Offline AI suite across phone and desktop

LLM Hub

Platforms:
Android, iOS
Price / license:
Freemium / source-available
Key difference:
Broader suite with agent, image, and translation features
Articles about LLM Hub (1)

Also mentioned in:

Competitor details change often; confirm each app's current price, license, and platforms on its own listing.

Compare in directory

Frequently Asked Questions

Who makes LM Playground?

Andriy Druk, an independent developer who publishes the app on Google Play under his own name and the code on GitHub since March 2024.

Is there an iPhone version?

No iPhone, desktop, or web version appears in the README, website, or Play listing; the project describes itself as an Android application.

Can I load my own models?

Yes. The README lists custom GGUF model files from any source alongside the built-in catalog; the formats beyond GGUF were not stated.

How much memory do I need?

There is no fixed minimum. The listing says the smallest model is 267 MB, 4B to 8B models benefit from 8 GB of RAM or more, and the 20B model needs a high-end device.

Does voice dictation work offline?

Per the website and release notes, yes after a one-time speech model download. It is not mentioned in the Play description, so check your installed version.

Where can I read the source?

The full code is in the LMPlayground repository on GitHub under the MIT license.

Verdict

LM Playground is a free, MIT-licensed way to run GGUF models on an Android phone: it combines a large built-in catalog with your own model files, on-device document Q&A, photo input, and dictation, and its public repository means its privacy claims can in principle be checked. Against that, it is Android-only, maintained by one developer, and its Play listing does not show the build number, so the version you install can differ from the newest GitHub release. It suits Android users who want an auditable, offline-first app; people on iPhone or desktop can compare PocketPal AI or Off Grid.

Sources

← Back to Power Local LLM

MCP