Talk naturally
Ask a question, continue a conversation, or request a concise answer. Your chat stays in the desktop app while you use local inference.
LOCAL-FIRST AI · WINDOWS DESKTOP
Useful AI.
Closer to you.
CapyRiver brings chat, code and everyday help into one calm workspace. Choose a compact 2B model or our adapted 4B option, then run inference locally on a compatible PC.
Two model sizes · local inference · transparent model origins
CapyRiverA thoughtful companion for your desktop01Designed for local use
02Two clearly documented model paths
03English · Русский · 简体中文
A DESKTOP ASSISTANT WITH A CLEAR JOB
Made to help you get things done.
CapyRiver is a Windows AI application for asking questions, drafting code and creating web content. The app manages model setup and gives you a focused place to work with a local assistant.
Build with code
Request source code, debugging help, Python, or a complete HTML page. State the output format and CapyRiver is designed to follow it.
See what fits
The Windows app can report CPU, memory, storage and supported NVIDIA telemetry so you can make an informed model choice.
Keep model choices understandable
See the model size, quantization, approximate storage needs and base-model provenance before installing. No crowded third-party catalog.
TWO WAYS TO RUN
Choose the fit.
Know what is inside.
CapyRiver is designed around practical, local model use. A smaller model uses fewer resources; a larger adapted model may handle more complex instructions, but still needs enough memory and can still make mistakes.
2BLIGHTWEIGHT OPTION
CapyRiver Core 2B
A compact 4-bit model choice for PCs with less available memory. It is based on upstream Qwen3.5-2B weights and is not a CapyRiver-trained model.
- QuantizationQ4_K_M
- Download size≈ 1.3 GB*
- ProvenanceUpstream base
4BADAPTED OPTION
CapyRiver Core 4B
A 4-bit NF4 base model with our CapyRiver LoRA adapter. The adapter applies project-specific tuning; it does not replace or erase the upstream model.
- QuantizationNF4
- Base download≈ 9.4 GB*
- ProvenanceQwen3.5-4B + CapyRiver LoRA
* Approximate figures; setup may require additional runtime files and storage. Model use is subject to the upstream licenses. We have not established that CapyRiver is generally better than other market models.
GET STARTED
Pick your device.
Follow the setup guide.
Download the Windows desktop app or install the Android app. Android downloads its 2B model on first use for offline chat.
WINDOWS · DESKTOP APP
Download EXECapyRiver for Windows
Windows 10/11 · x64 · local chat and model setup
ANDROID · ON-DEVICE APP
Download APKCapyRiver for Android
Download the 2B model once, then chat locally on ARM64 Android phones.
Android 7.0+ · ARM64 · 1.22 GiB model download required on first use
IPHONE / IPAD · IPA GUIDE
Install an IPA with a cable
An IPA package is not published yet. These are the steps to install a signed build once one is available.
01
Use a Mac with Xcode to build and sign the iOS app. An IPA must be signed for App Store, TestFlight, or registered-device distribution.
02
Connect iPhone to the Mac using a data-capable USB cable. Tap Trust on the iPhone and allow the Mac accessory prompt if it appears.
03
For a development install, select the iPhone in Xcode and run the app. For an Ad Hoc IPA, use Apple Configurator and a provisioning profile that includes your device.
04
Enable Developer Mode if iOS asks, restart, then confirm. Windows can recognize and sync an iPhone over USB, but it cannot sign an IPA or install this app by simply copying the file.
CapyRiver iOS app / signed IPA: not available in this release.Apple’s official cable-install guide
WHY CAPYRIVER
A different set
of trade-offs.
We focus on local setup, model transparency and a compact desktop workflow. “Better” depends on what you need: CapyRiver is not claimed to outperform every commercial or open model.
Local by design. Once the model is installed, inference runs on your compatible PC; the app does not route prompts through a CapyRiver cloud chat service.
Smaller choices. The 2B option aims for lower resource requirements, with a 4B adapter path for PCs that can accommodate a larger model.
Clear provenance. We identify the upstream bases and adapter so users can inspect what they download and which licenses apply.
GOOD TO KNOW
Questions, answered.
Details about models, privacy, installation and the iPhone preview.
Is CapyRiver smarter than other models?
We do not make a general superiority claim. Model quality varies by task, prompt, hardware and settings. The current release does not include an independent benchmark that proves it is better than the wider market.
What is the difference between the 2B and 4B options?
The 2B option uses quantized upstream base weights and is intended to use fewer resources. The 4B option adds a CapyRiver LoRA adapter to an upstream 4B base and has substantially higher memory and storage needs. Both can make mistakes.
Does the Android APK run AI?
Yes. The Android app downloads the upstream Qwen3.5-2B GGUF model on first use, then runs chat locally. The model is about 1.22 GiB and needs available phone memory.
Can I download the iPhone IPA now?
There is no signed CapyRiver IPA in this release. A real iOS build must be compiled and signed with Apple’s tooling and an appropriate distribution profile. The guide above explains the cable install route when a compatible IPA exists.
Does local mean no internet is ever used?
Model and runtime downloads need internet. Normal inference can run locally after setup. Optional external services or web features send the relevant request to those services; review each connection before using it.
What are the licenses?
The app source, upstream model weights, adapter, runtimes and bundled tools may use different licenses. Check the project’s LICENSES and third-party notices before redistributing anything.