Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM/TokForge Review: Offline Android AI Chat with Roleplay and Image Generation
Mobile & Edge LLMs

TokForge Review: Offline Android AI Chat with Roleplay and Image Generation

·9 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

Founder-reviewed. TokForge's founder reviewed this entry's technical specs and description for accuracy. The review itself remains independent PromptQuorum editorial content.

TokForge (version 1.3.6.1, confirmed by the developer on 9 October 2026) is a free Android app by Defcon-One that runs local language models through llama.cpp or MNN and adds roleplay characters, on-device image generation, voice, and document Q&A, with no account. Its listing says conversations never leave the device, that web search is off by default, and that a benchmark can optionally post results to a public leaderboard. The app is proprietary closed-source software, the developer states it has no ads, in-app purchases, or subscriptions, and the iPhone and iPad version is a public TestFlight beta that is not yet on the App Store.

TokForge is a free Android app, published by a developer listed on Google Play as Defcon-One, that runs language models on the phone and combines them with roleplay characters, on-device image generation, text-to-speech with voice cloning, document Q&A, and a built-in benchmark that can post speeds to a public leaderboard. Version 1.3.6.1 is a production release on Google Play, while the iPhone and iPad version is a public TestFlight beta that anyone can join. The app is proprietary, closed-source software and its source code is not published. This review is based on the Google Play listing and the developer's own website at tokforge.ai, checked on 3 October 2026; PromptQuorum has not tested the app hands-on.

TokForge Review: Offline Android AI Chat with Roleplay and Image Generation

Key Takeaways

  • Developer: listed as Defcon-One on Google Play, with Isaac Maple and a United States location shown in the developer details and a contact email on the listing.
  • Price and license: free with no ads, in-app purchases, subscriptions, or account according to the developer's FAQ; proprietary and closed source, with the license in section 2 of the terms.
  • Engines: llama.cpp (GGUF) with OpenCL and Vulkan paths, MNN with OpenCL and Vulkan (which one is used depends on the phone), or an OpenAI-compatible server you connect yourself.
  • Scope: roleplay characters, on-device images, Kokoro voices with cloning, document Q&A, a 52-model catalog with Hugging Face search, and an optional public speed leaderboard.
  • Signals: version 1.3.6.1 as confirmed by the developer on 9 October 2026, 5K+ Google Play downloads, and a listing updated on 20 September 2026 (checked 3 October 2026).

📍 In One Sentence

TokForge is a free, closed-source Android app by Defcon-One that runs local language models with roleplay characters, on-device image generation, voice, and document Q&A, with its iPhone and iPad version still in TestFlight beta.

💬 In Plain Terms

You install it from Google Play, download a model that fits your phone, and chat, make images, or talk to characters without an account; the developer also runs a public leaderboard where you can choose to post your phone's speed results.

📌Note: This review is based on the Google Play listing and the developer's own website, checked on 3 October 2026. Performance figures on the site are the developer's claims, and PromptQuorum has not tested or benchmarked the app.

What Is TokForge?

TokForge is an offline-first Android chat app that bundles a local model runner with roleplay, image, voice, and document tools. According to its Google Play listing, it runs full models on the phone with no cloud or account, supports characters imported from chub.ai and TavernAI cards, and picks the fastest engine for your chip automatically.

The developer says the app is much smaller than established alternatives. That candor matters for expectations: it has a small install base, and its iPhone and iPad version is only in public TestFlight beta, not yet on the App Store.

Get It

TokForge is distributed through Google Play on Android; the Apple version is a TestFlight beta rather than an App Store release.

Platform
Where to get it
AndroidGoogle Play
iPhone / iPadPublic beta via TestFlight (see tokforge.ai)
Source codeNot published

Version 1.3.6.1, confirmed by the developer on 9 October 2026; older pages and listings may still show 1.0.

How to Get Started

The sources describe the flow only in outline; PromptQuorum has not run these steps.

  1. 1
    Install and let the app size up your phone
    Why it matters: The app detects device capabilities and curates the models offered, according to the developer's website.
  2. 2
    Download a model
    Why it matters: Choose from the curated catalog of 52 models or search Hugging Face through the built-in downloader.
  3. 3
    Optionally run the benchmark
    Why it matters: AutoForge benchmarks model and backend combinations on your hardware; posting to the public leaderboard is your choice.
  4. 4
    Chat, create, or ask your documents
    Why it matters: Start a chat, create or import a character, generate an image, or add a document and ask about it.

Features Confirmed by the Sources

Every item below comes from the Google Play description or the developer's website; none has been independently tested.

  • Chat and roleplay. Create characters or import cards from chub.ai and TavernAI (PNG or JSON), with personas, lorebooks, alternate greetings, and group chats.
  • Images. On-device generation that runs on the CPU everywhere, with GPU acceleration on Adreno and NPU acceleration on supported Snapdragon devices. The developer cites about 31 seconds for a 512x512 image on a Pixel 9 Pro XL.
  • Voice. Kokoro text-to-speech with several voices, voice cloning from about a minute of recorded audio, and voice input.
  • Documents and memory. Questions about PDF, DOCX, EPUB, Markdown, CSV, and text files, plus long-term memory and a knowledge graph.
  • Engines and models. llama.cpp (GGUF) with OpenCL and Vulkan, MNN with OpenCL and Vulkan depending on the phone, or an OpenAI-compatible server, with automatic routing by chip and a 52-model catalog.
  • Developer API and backup. Local API endpoints for automation (270+ per Google Play, 284 per the website) and backup and restore of conversations, characters, and settings.

Hardware and Speed

The developer's website states a minimum of 4 GB of RAM for small models and 8 GB or more for larger ones. Speed depends on the chip: the listing describes GPU paths on Adreno and NPU paths on supported Snapdragon devices, with CPU fallback elsewhere.

The app's AutoForge benchmark measures each model and backend combination on your phone, and the public leaderboard at leaderboard.tokforge.ai collects device-submitted results if you choose to post. Those figures are user-submitted and developer-hosted, not independent measurements.

Privacy and Online Features

The Google Play Data safety section declares "No data collected" and "No data shared with third parties", and the description claims zero analytics and zero telemetry. The website adds that chats are stored only on the device.

Three features go beyond that picture, by the listing's own description: web search is off by default and, when turned on, fetches current information; posting results to the public leaderboard is optional; and a connected OpenAI-compatible server receives whatever you send it. Because the source is not published, none of this can be checked against code, and these are the developer's declarations, not audit results.

📌Note: PromptQuorum has not inspected the app's network traffic. Anyone handling confidential data should verify behavior with web search and leaderboard posting off and on.

From the Maker

Isaac Maple, the developer behind TokForge, shared the following about the app and the reasons for building it. It is presented as the developer's own words, lightly edited for readability, not as PromptQuorum's independent editorial assessment:

"I think our best feature is autoForge, which finds the best and fastest inference setup for each handset and model: CPU, OpenCL or Vulkan, thread count and context size.

The app is really just me and one other buddy, guardian37x in the Discord.

I convert and try to optimize a lot of the smaller LLMs for Edge AI and upload them to my Hugging Face account. I think there are around 85 models there now.

As for why I built it: when I started, there were a few other options for mobile inference, but most were llama/GGUF based, cost money, were filtered, had ads, and weren't really performance-based or didn't have an API backend to really dive in. I also hadn't seen any that used MNN, and as an engineer, MNN at that time was sometimes 50% faster than GGUF, which was really exciting.

So that's why I built it. I figure AI inference costs will keep going up and become more expensive for regular people to access AI they own, on their own devices, that isn't filtered or censored. So I wanted to make something that addressed that and was fast, free and without ads."

— Isaac Maple, developer

Trade-Offs: Benefits vs. Limitations

Free and no account

What it means in real use:
Install and use without sign-up or subscription.
Limitation / caveat:
Proprietary and closed source; no public repository.

Roleplay and media in one app

What it means in real use:
Characters, images, voices, and documents sit in one install.
Limitation / caveat:
Breadth is described by the developer; quality is untested here.

Chip-aware engines

What it means in real use:
llama.cpp, MNN, and a remote server with automatic routing.
Limitation / caveat:
Best paths depend on the SoC; speeds vary a lot by phone.

Public benchmark data

What it means in real use:
A leaderboard shows how phones perform before you commit.
Limitation / caveat:
Results are user-submitted and hosted by the developer.

Android release is stable-labeled

What it means in real use:
Version 1.3.6.1 is a normal production release on Google Play.
Limitation / caveat:
Small install base (5K+), and the iPhone version is a beta.

Who Should Use It

  • Android users who want offline roleplay with imported character cards. Chub.ai and TavernAI card import is a headline feature.
  • People who want images, voice, and documents without cloud services. All three are described as running on the device.
  • Users who like seeing benchmark data for their chip. The built-in benchmark and leaderboard are designed for that comparison.

What We Could Not Verify

  • Source code. TokForge is proprietary and closed source, with the license in section 2 of its terms, so behavior cannot be checked against code.
  • Hands-on performance and quality. PromptQuorum did not run the app, so speed, battery use, and output quality are not assessed.
  • The no-ads claim. The developer's FAQ says there are no ads, in-app purchases, or subscriptions; PromptQuorum did not install the app to confirm this.
  • Developer background. Google Play names Defcon-One, with Isaac Maple and a United States location in the developer details; no company registration or track record was found in the sources read.
  • iPhone and iPad. The public TestFlight beta, open to anyone, was not tested, and the app is not yet on the App Store.

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)

+20 more not shown

Layla

Platforms:
Android, iOS
Price / license:
Paid / closed source
Key difference:
Companion and roleplay focus with an optional cloud mode
Articles about Layla (4)

Also mentioned in:

LLM Hub

Platforms:
Android, iOS
Price / license:
Free + IAP / PolyForm NC
Key difference:
Wider on-device suite with an MCP agent and media generation
Articles about LLM Hub (1)

Also mentioned in:

Private Mind

Platforms:
iOS, Android
Price / license:
Free / MIT
Key difference:
Open-source offline chat with on-device document Q&A
Articles about Private Mind (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

Is TokForge free?

Yes. Google Play and the developer's website describe it as free with no account, and the developer's FAQ states there are no ads, in-app purchases, or subscriptions.

Is TokForge open source?

No. TokForge is proprietary, closed-source software; the license is in section 2 of its terms, and no public source repository exists.

Who makes TokForge?

Google Play lists the developer as Defcon-One, with Isaac Maple and a United States location in the developer details and a contact email on the listing.

Does it work offline?

Per the listing, chat, images, and voice run on the device once models are downloaded. Web search is off by default, leaderboard posting is optional, and downloads and any connected server need a network.

Is there an iPhone version?

Yes, as a public beta on TestFlight that anyone can join; it is not on the App Store yet. The Android version is on Google Play.

What hardware does it need?

The developer's website states at least 4 GB of RAM for small models and 8 GB or more for larger ones, with GPU and NPU paths on supported Snapdragon phones.

Which engines and models does it use?

llama.cpp for GGUF models, MNN, or an OpenAI-compatible server you connect yourself, plus a 52-model catalog and a Hugging Face downloader.

Can it import character cards?

Yes. The description says cards from chub.ai and TavernAI (PNG or JSON) can be imported, with personas, lorebooks, alternate greetings, and group chats.

How does it compare with PocketPal AI?

PocketPal AI is a free MIT-licensed chat client on iOS and Android, while TokForge is closed source with roleplay, image, and voice features on Android and an iPhone beta.

Verdict

TokForge packs roleplay characters, on-device images, voice cloning, document Q&A, and chip-aware engine routing into a free Android app that, per the developer, has no ads, in-app purchases, or subscriptions, and its public benchmark leaderboard is an unusual attempt at transparency about speed. Against that, the software is proprietary and closed source, the install base is small at 5K+, the iPhone and iPad version is a public TestFlight beta that is not yet on the App Store, performance figures come from the developer, and nothing here has been tested hands-on. It suits Android users who want a free offline roleplay-and-media app and accept those terms; readers who want auditable code can compare PocketPal AI or Private Mind.

Sources

  • TokForge on Google Play — description, developer details, Data safety section, download count, and last-updated date, checked 3 October 2026.
  • tokforge.ai — version, platforms, hardware guidance, and the developer's own comparison guide, checked 3 October 2026.
  • TokForge terms — license (section 2): proprietary software, checked 9 October 2026.
  • TokForge FAQ — free, with no ads, in-app purchases, or subscriptions, and the public TestFlight beta, checked 9 October 2026.

← Back to Power Local LLM