Strata

1. What is Strata?

Strata is a free, open-source (MIT licence) program that lets you run a very large AI model on your own gaming PC, with no server to rent and no monthly subscription. The model it runs is Qwen3.8-Flash-Next from the Qwen team. According to the project, it has about 125 billion parameters, which normally calls for an expensive server. Strata shares the work between your graphics card, RAM, processor and SSD so that it fits on a home computer.

Once it is running you get a chat window in your browser (at http://127.0.0.1:8080, reachable only from your own PC). You can ask questions, have it write text or code, or have it read pictures if you enable that option. The project says everything runs on your machine and nothing leaves it. Your chats are stored only in your browser.

Good for you if:

• You have a desktop or gaming laptop with a fairly strong dedicated graphics card (see section 2) and plenty of RAM.
• You want a capable AI whose data stays on your own machine, or you do not want to pay a monthly AI subscription.
• You do not mind a one-time download of roughly 70 GB and a few minutes of waiting at each start.

Not for you if:

• You have an office laptop, a PC with only integrated graphics, or a card with less than 12 GB of video memory (VRAM). The documentation says integrated AMD graphics are not supported. An 8 GB NVIDIA card is described as working but slowly.
• Your PC has less than 32 GB of RAM.
• You use macOS. The documentation covers Windows and Linux only.
• You want something as light and instant as a phone chatbot. This is a heavy program that loads the model each time you start it.

Note: The speed figures in the project's documentation were measured by its authors on their own PCs (for example an RTX 5070 with 12 GB and 64 GB of RAM). Yours may be faster or slower.

2. What do you need?

According to the official documentation:

Part	Requirement
Graphics card	NVIDIA GeForce RTX 20, 30, 40 or 50 series; or AMD Radeon RX 7900 XT/XTX, RX 7800 XT / 7700 XT, RX 9060 XT, RX 9070 / 9070 XT, Radeon AI PRO R9700, RX 6800 / 6900. It needs 12 GB of VRAM or more
RAM	32 GB or more. 64 GB runs every model size
Disk	About 80 GB free. Use an SSD, ideally NVMe: the first start is much faster
System	Windows 10 / 11 or Linux
Graphics driver	NVIDIA: version 580 or newer (https://www.nvidia.com/drivers or the NVIDIA App). AMD on Windows: a current AMD Software Adrenalin Edition driver from https://www.amd.com/en/support
Processor	x86-64 with AVX2 (most Intel/AMD CPUs from roughly the last 8 years)

You do not need to install Python or any developer tools yourself. The documentation says the installer sets up Python, the engine that runs the model, and the model. The only thing you do beforehand is update your graphics driver.

No account, no API key, no cost. The internet is needed only to download the program and the model the first time.

To check your PC on Windows: right-click the taskbar, open Task Manager, and look at the Performance tab. It shows your RAM and GPU; under GPU, "Dedicated GPU memory" is the VRAM.

3. Step-by-step install

These steps are for Windows; Linux has an equivalent at the end of this section. The latest release at the time of writing is v0.1.41 (8 Oct 2026); the project updates very often.

Step 1: Update your graphics driver

Update the NVIDIA driver (version 580 or newer) or AMD Adrenalin using the links in section 2, and restart if the driver installer asks. If the NVIDIA driver is too old, Strata will say so when you run it.

Step 2: Download Strata

Open https://github.com/Niko1221/Strata, click the green Code button and choose Download ZIP. Or download directly: https://github.com/Niko1221/Strata/archive/refs/heads/main.zip

This ZIP holds the setup scripts, not the AI model itself (the model is downloaded in a later step). Download only from this Niko1221/Strata page and avoid look-alike sites.

Step 3: Unzip it

Right-click the ZIP, choose Extract All… and pick a folder you will remember, such as D:\Strata. Choose a drive with plenty of free space: the model is stored in a Strata-data folder next to the Strata folder and takes 70 to 120 GB.

You should see files such as START-HERE.bat and UPDATE.bat.

Step 4: Run the installer

Double-click START-HERE.bat. A black command window opens and asks a few questions:

• Which model and which size.
• How much text the model keeps in mind at once (called "context").
• Whether it should read pictures.

Just press Enter for each question to accept the recommended choice for your PC. The installer detects whether you have an NVIDIA or AMD card and picks the right engine.

If Windows blocks a file because it is unsigned (Smart App Control), see section 5 before continuing.

Next it downloads the model, about 70 GB for the smaller sizes (the larger ones go beyond 90 GB). How long that takes depends on your connection. If it is interrupted, run START-HERE.bat again and it continues where it stopped.

Step 5: Wait for the model to start

When the download finishes, Strata starts the model. For 1 to 3 minutes (longer the first time) your PC may be slow or freeze. The documentation explains that Strata loads 35 to 55 GB into RAM. This is normal. Do not close the window. It shows what Strata is doing.

When it is done, your browser opens Strata at http://127.0.0.1:8080. If you see a page with Chat, Monitor and About tabs, it worked.

Step 6 (optional): Linux

On Linux, run this inside the Strata folder. Ubuntu 22.04 and 24.04 get everything installed automatically.

./setup.sh

4. Your first run

1. In your browser, open http://127.0.0.1:8080 if it did not open by itself.
2. On the Chat tab, type a short request, for example "Write a two-day leave request email." The answer appears word by word.
3. The chat has a thinking menu: off, low, medium, high. Off is fastest; high suits hard questions.
4. To let the model read pictures, you must have answered yes to the picture question during setup. Then click Picture in the chat. On AMD cards under Windows, the documentation says picture reading does not work yet.
5. The Monitor tab shows the model and your GPU, CPU and RAM use live.
6. Next time, double-click START-HERE.bat. The documentation says the model loads in about 30 to 90 seconds and nothing is downloaded again. Closing the command window stops the model.

If you use a coding assistant or another app, you can point it at http://127.0.0.1:8080/v1 (an "OpenAI-compatible" provider); any API key works. This is for technical users and you can ignore it.

Security: By default Strata is open only on your own PC. Do not open it to your phone or other PCs (the --host 0.0.0.0 option) without setting a key. The documentation says to always set a key when you do.

5. Common problems and fixes

The PC freezes or crawls the first time it starts Cause: Strata is loading 35 to 55 GB into RAM. Fix: wait, and do not close the window. If it is still frozen after 10 minutes, restart the PC, close other programs (browsers use a lot of RAM) and try again. If it keeps happening, run SETUP.bat and choose a smaller size (Q2_0 or IQ2_XS).

It says the NVIDIA driver is too old Update the driver (NVIDIA App or https://www.nvidia.com/drivers), restart, and run START-HERE.bat again.

"Windows blocked the Strata engine… Smart App Control" On a clean Windows 11, Smart App Control refuses unsigned programs. The documentation says the project's strata.exe and strata-vision.exe are not signed. The official fix is to turn Smart App Control off in Windows Security > App & browser control > Smart App Control settings. Think first: the documentation notes it cannot be turned back on without resetting Windows. If only the picture reader is blocked, you can run START-HERE.bat --setup --vision no to disable picture reading and keep Smart App Control on.

"Port 8080 is already in use" Usually Strata is already running, so look for its window. If another program uses the port, run START-HERE.bat --port 8081.

Very slow, the disk light blinks constantly, or "the engine stopped unexpectedly" Your PC is out of free RAM. Close other programs or choose a smaller size. After "engine stopped", simply send your message again; Strata restarts the engine itself.

It says the prompt exceeds the context The conversation is longer than the context you chose. Start a new chat, or run SETUP.bat and pick more context.

The download stopped halfway Run START-HERE.bat again; it continues where it stopped.

Still stuck? See the project's troubleshooting page: https://github.com/Niko1221/Strata/blob/main/docs/TROUBLESHOOTING.md

6. Uninstall and update

Update: close the Strata window, then double-click UPDATE.bat. It updates the program without starting the model and without downloading the model again. If you got Strata as a ZIP (not with Git), the documentation says UPDATE.bat cannot fetch new files itself: download a fresh ZIP, unzip it anywhere and run START-HERE.bat. The model lives in the Strata-data folder next to the Strata folder, so the new copy finds it with no re-download.

Uninstall: Strata has no uninstaller. Close its window, then delete two things: the Strata folder (the program, .venv, engine) and the Strata-data folder beside it (the 70 to 120 GB model). The documentation records the data folder's location in %APPDATA%\Strata\settings.json; delete the %APPDATA%\Strata folder too for a clean sweep. Your chats live in the browser, so clear the site data for 127.0.0.1:8080 in the browser settings to remove them.

7. Frequently asked questions

Does it cost anything? No. Strata and the model are provided free. The author has an optional donation button (Buy Me A Coffee). Strata is MIT-licensed; a few parts and every model have their own licences, listed at https://github.com/Niko1221/Strata/blob/main/docs/HOW_IT_WORKS.md#license if you plan to use it commercially.

Does it need internet? Only to download the program, engine and model the first time. After that, chatting runs on your PC.

Is my data sent anywhere? According to the project, no: nothing leaves your PC. That is the project's own statement; I did not verify it by monitoring network traffic.

Will it run on a weak PC? Probably not. It needs a dedicated card with 12 GB of VRAM or more and 32 GB of RAM. The documentation lists experimental setups for older cards and CPUs, but they are slow and community-tested.

Does it run on a Mac? The documentation supports Windows and Linux only.