Install Strata on Windows or Linux and send a first prompt

Strata 0.1.43 runs on Windows 10/11 or Linux. The GPU baseline is an NVIDIA RTX 20/30/40/50-series card or a Radeon family listed in the documentation. NVIDIA driver 580 or newer is recommended. AMD uses a current Adrenalin driver on Windows and the kernel's amdgpu driver on Linux. Check your GPU and driver in Windows Device Manager, AMD Software, or Linux device information.

RAM needs depend on the model size; 32 GB is a starting point to consider. With 64 GB, every model size in the guide can run, though large variants and long contexts still need room for other apps. Reserve about 80 GB for the roughly 70 GB model, the roughly 6 GB MTP⁵ layer, settings, and possible image files. On Linux AMD, setup may need about 10 GB more for ROCm⁶. An NVMe⁷ SSD⁸ helps reduce the first model load time.

This guide installs Strata 0.1.43. The Git command below checks out that fixed version. Without Git, download and extract the ZIP from the same v0.1.43 release. The update instructions below distinguish the installation methods. For a new installation, you can choose the 0.1.43.1 ZIP with the Linux NVIDIA regression fix and follow the same steps below.

From the repository folder, double-click START-HERE.bat on Windows. On Linux, open a terminal in that folder and run ./setup.sh. If Linux reports a permission error, start it with bash setup.sh. Setup prepares the required Python environment, GPU engine, model, and MTP draft layer.

If the NVIDIA driver is older than 580, update it before setup. If no ready-made NVIDIA engine matches the PC, setup explains the build-tool installation before compiling an engine. AMD Windows uses the Adrenalin driver and bundled HIP⁹ engine; AMD Linux also prepares ROCm and compiles an engine for the card.

The setup questions determine model size and context for your available RAM and GPU memory. Keeping the recommended choices continues setup. If you will not send images, disable the vision option to leave more GPU memory for the expert cache. Longer contexts also use more KV-cache¹⁰ memory, so start with the length you need.

Setup downloads the roughly 70 GB model and runtime files. If a first-install download is interrupted, rerun START-HERE.bat on Windows or ./setup.sh from the repository folder on Linux. Setup resumes with files already downloaded. During the first load into RAM and GPU memory, the PC may slow down or appear unresponsive for 1–3 minutes. Keep the window open while progress messages continue.

When setup finishes, the browser opens or the local address appears in the window. Enter a short prompt in Chat at http://127.0.0.1:8080 and check that an answer appears. The server window shows that the model has started. Later runs do not download the same files again. Close the server window to stop it.

For the fixed 0.1.43 checkout above or a ZIP installation, move to a new release ZIP. Keep the old Strata folder and extract the new ZIP under the same parent folder, then run START-HERE.bat or ./setup.sh. Finding the adjacent existing Strata-data folder avoids downloading the large model again. Settings live inside the old Strata folder, so check the model, context and working folder in the new setup.

A branch checkout created with git clone without pinning a version can update through Windows UPDATE.bat or Linux ./update.sh in that folder. After the script refreshes code, engine, packages and settings, start the model with START-HERE.bat or ./setup.sh. In the fixed-tag checkout above, the updater's git pull fails; use the new-ZIP route instead.

If the first download does not finish, check the network connection and free disk space, then rerun START-HERE.bat on Windows or ./setup.sh on Linux. Setup resumes with files already downloaded. Use UPDATE.bat or ./update.sh only when updating a branch-based Git installation; use a new release ZIP for fixed-tag or ZIP installs. If setup says the NVIDIA driver is old, update it, restart the PC, and run the launcher again.

If you see port 8080 is already in use, first check whether a Strata server window is already running. If another program owns the port, consult the installation guide for the port option and start Strata on another port. If the server is running but the browser did not open, enter the local address yourself.

If model loading stops or you see engine stopped unexpectedly, restart the PC, close memory-heavy apps such as browsers, games, or editors, and try again. If it still stops, check free RAM and VRAM¹¹ and choose a smaller model such as Q2_0 or IQ2_XS, or a shorter context. If Linux AMD fails while installing ROCm or compiling, note the final terminal error and selected GPU architecture, then consult the official AMD HIP guide.

Once the first answer appears, the installation flow is complete. If the reply is slow or stops, check GPU, CPU¹², and RAM use in Monitor. The stage of failure points to the next check: model size, driver and runtime match, or whether the download completed.

This is for an existing low-memory Windows setup that is unusually slow, not part of the first installation. When GGUF¹³ shards remained mmaped during unbuffered NTFS reads, expert files could be read one at a time. Strata 0.1.43 closes each shard's mapped view when the file tier begins unbuffered reads, restoring parallel expert-file reads.

To compare after updating, use the same GGUF model and prompt. The release example used IQ3_XXS on an RTX 5070, Ryzen 5 7600, and 64 GB RAM. Short-prompt decode¹⁴ changed from 13.3 to 18.2 tok/s with 22 GiB of experts held in RAM, and from 4.6 to 8.5 tok/s with 12 GiB. These are results for that Windows setup.