Strata: Run Alibaba's 125B Qwen AI on Your Own PC for Free (Setup Guide)

Most powerful AI models run on someone else's server. Your prompts travel to a data center, get processed, and come back. That works - but it means your conversations, your code snippets, and the files you share are all leaving your machine.

Strata is a free, open-source app built by a developer named Niko1221 that runs a large AI model entirely on your own PC. The model it uses is Qwen3.8-Flash-Next - a 125-billion-parameter model from Alibaba's Qwen team, released on HuggingFace. This guide covers what Strata is, whether your PC can handle it, and how to get it running.

What Is Qwen3.8-Flash-Next?

Qwen3.8-Flash-Next is a 125-billion-parameter language model built by Alibaba's Qwen team and released on HuggingFace. "125 billion parameters" is a rough measure of the model's size and complexity - larger models generally handle more nuanced questions and longer context.

To put the number in perspective: the public versions of GPT-3 had 175B parameters, and GPT-4's architecture has never been officially disclosed. A 125B model in this class is genuinely large. The important point for this guide is not benchmark comparisons - those shift constantly. What matters here is that this is a capable, serious model that you can now run locally, for free, with no account required.

What Is Strata?

Strata is a free, MIT-licensed open-source app and inference engine built by Niko1221. It had over 28,000 GitHub stars as of October 19, 2026, and is on version 0.2.0-beta - actively developed but functional. The MIT license means you can use it, modify it, and build on it without restrictions.

Here is what Strata does on top of just running the model:

• Chat interface - a clean local web UI to talk to the model
• Code writing - the model can write and explain code
• Image reading - image input is on by default; paste or upload a picture and ask questions about it
• OpenAI and Anthropic API compatibility - Strata speaks both API formats on localhost, so tools like Claude Code, Codex CLI, and OpenCode can connect to it as if it were a cloud service
• Private by design - no account, no cloud sync, no telemetry; your chats and files never leave your machine

Credit fully to Niko1221 for building this. The repo README is clear and well-maintained. If you find it useful, give it a star.

Can Your PC Run It?

This is the first question to answer before downloading anything. The model is large, and Strata is honest about what it needs.

Minimum requirement: at least 32 GB of combined RAM and graphics memory (VRAM). For example, a PC with 16 GB of RAM and a GPU with 16 GB VRAM meets the minimum. A 32 GB RAM machine with a 12–16 GB GPU card runs the Coder model size well. 64 GB of RAM runs every model size Strata offers.

Disk space: about 80 GB of free storage. SSD is recommended for faster model load times.

Operating systems: Windows 10/11, Linux, or macOS.

Supported GPUs:

• NVIDIA RTX 20, 30, 40, or 50 series (12 GB VRAM or more)
• AMD Radeon RX 6800 and newer
• Intel Arc
• Apple Silicon Mac (M1 and later)
• AMD Ryzen AI Max

If your machine does not meet the 32 GB combined threshold, the model will either not run or will run very slowly. Strata's in-app screen reads your hardware and recommends the model size that fits - so you are not guessing.

Install Option A - The Strata App (Easiest)

This is the recommended route for most people. No terminal required.

Step 1: Go to the Strata releases page and download the file for your system and GPU:
github.com/Niko1221/Strata/releases/latest

Step 2 - Windows: Run the downloaded .exe. No admin rights needed. If Windows shows "Windows protected your PC", click More info then Run anyway - the installer is not yet code-signed, but it is open source and you can read every line on GitHub.

Step 2 - Linux: Make the AppImage executable and run it:

If you are on Ubuntu 22.04, you may need FUSE 2 first: sudo apt install libfuse2. On Ubuntu 24.04 or newer: sudo apt install libfuse2t64.

Step 2 - Mac: Open the downloaded .dmg and drag Strata to Applications.

Step 3: On the first screen Strata scans your PC and recommends a model size that fits your hardware. Press Install. The model download is large - roughly 40 to 70 GB depending on the size - but it resumes automatically if the connection drops or you close the window. Closing the window also stops Strata from running.

Install Option B - One-Line Terminal Install

Better for servers, automated setups, or if you want to connect AI coding agents to Strata. The script checks your hardware, downloads the right files, and verifies the SHA-256 hash. No admin rights needed.

Linux / Mac:

Windows PowerShell:

You can also paste the GitHub repo link into any AI coding tool and ask it to read the README and follow the install steps - that works too.

Connect Claude Code, Codex CLI, or OpenCode

Once Strata is running, it exposes a local API endpoint. Because it speaks both the OpenAI and Anthropic API formats, you can point tools like Claude Code, Codex CLI, and OpenCode to your localhost address instead of the cloud.

Check the Strata README for the exact localhost port and config snippet for each tool - it is documented there and may update across versions. The general idea is: set the API base URL to your local Strata endpoint, set a dummy API key (since no auth is needed locally), and the tool connects as if it were hitting a real cloud API.

Honest Expectations

A few things worth being clear about before you start:

• Speed depends on your hardware. On the minimum 32 GB setup, responses will be slower than a cloud API. On a 64 GB machine with a strong GPU, it is noticeably faster.
• The first download is large. 40–70 GB takes time. Plan for it.
• This is beta software. Version 0.2.0-beta means things can change or break. Keep an eye on the GitHub releases page.
• No benchmark claims here. The reel this post accompanies makes a "beats ChatGPT and Claude" claim - that is not verified and we are not repeating it. Qwen3.8-Flash-Next is a strong, capable model. Whether it outperforms specific closed models on your specific tasks is something you will need to test yourself.
• Privacy is real. Nothing leaves your machine. No account, no analytics, no telemetry. That alone is worth something.