Strata-AMD
Strata AMD runs large local AI models, such as the Qwen3.8-Flash-Next mixture-of-experts model, on an AMD GPU plus system RAM and serves them through an…
Platform availability
Available on 4 of 10 platforms
Health score?A 0-100 score computed nightly from four factors: maintained (40%: recent commits and steady activity), popular (25%: GitHub stars, log scale), easy to install (20%: platforms, install notes, setup guides, Docker image) and light to run (15%: minimum RAM, ARM support). When the source code has been checked, a fifth factor, security check (20%: automated scan for committed secrets, unsafe desktop-app settings, known-vulnerable dependencies and project hygiene; any serious finding caps it at 20), is added. Factors without data are left out and the rest are rescaled.
72/100 Good
- Maintained
- 86last commit 15 days ago
- Popular
- 9240,996 GitHub stars
- Easy to install
- 654 install platforms, install notes, Docker image
- Light to run
- 1032 GB minimum RAM
Availability: platform listings confirmed today
- GitHub stars
- 40,996
- Open issues
- 1045
- Last commit
- 2026-11-06
- Latest release
- v0.2.11
- License
- MIT
- Activity
- rising
Checked 15 days ago - source: GitHub (Niko1221/Strata)
Resources & Compatibility
Minimum requirements: 32 GB RAM.
Checked 15 days ago - source: Docker Hub image tags and published docs
First-install notes
- Run setup first and pick your model Run ./setup.sh (Linux/macOS) or START-HERE.bat (Windows) before starting the app, and choose the model pack, context size (how much conversation it can remember), and vision on/off. These choices decide whether the model fits your memory and what it can do, and setup saves them so later runs start the already-configured model. source
- Note the local web address and port After setup, the server runs locally at http://127.0.0.1:8080. That address is for your own machine, so keep it available if you plan to open the app or connect local tools there. source
- Keep your generated config/profile Setup also creates a model profile/config file that records the downloaded model and your options. Keep that file safe because tools like strata doctor and strata serve use it to inspect or start the model later. source
- Right-size the memory budget on shared servers By default Strata copies the model's experts into system RAM even when the GPU already holds them, so on a server running other containers the process can take most of your memory (measured: about 59 GiB for the container, leaving 4 GiB for everything else). On a machine with two 32 GB cards that also runs other workloads, set Engine settings to resident-budget-gib=8 and Parallel requests to 2; fo source
- AMD graphics setup on Unraid You need the amdgpu kernel driver, which Unraid ships as standard on Linux 6.8 and newer, so no driver plugin is required. When configuring, pick the engine preset that matches your AMD card family. source
- Desktop install On macOS, Windows, and Linux, download the latest build from the official GitHub releases page before running setup. source
Fetched yesterday - each note links to its source
Variants
Alternatives
No alternatives collected yet.
Common questions
Which Qwen3.8-Flash-Next quantized packs are documented for Strata?
Strata lists ISTA-DASLab as the recommended provider, with Q2_0, IQ2_XS, IQ3_XXS and IQ3_S choices and IQ1_M noted as the coder option. Unsloth also provides several quantized packs, while Gyro offers lower-memory but slower options.
My terminal cannot find the Strata command after installation. How do I fix it?
On Windows, open a new PowerShell window. On Linux or Mac, use the full installer path, normally ~/.local/bin/strata, and add that directory to PATH if needed before running strata --version.
Can Strata be used with dual AMD GPUs on Windows?
Community reports say it can run on Windows with dual AMD GPUs despite no official support. The sourced result reports about 1350 pp on a 9070 XT system.
What happens if a model does not fit in RAM?
The parts that do not fit in memory are read from the SSD. This means the setup needs a lot of free disk space.
What performance was measured with one Radeon AI PRO R9700 and 251 GB RAM?
The source measured 21 to 25 tokens/s output and 320 to 360 tokens/s prompt on an 8K context.
Answers sourced from agihunt.info, github.com, github.com
Community
Collected 5 days ago - public community threads