Strata is a local inference engine for developers and PC users to run Qwen models and connect them to their apps.
A local model that uses more than your graphics card
The Niko1221/Strata repository is the original project behind the Strata Qwen searches. It runs Qwen3.8-Flash-Next through a local browser interface and API. This is an inference runtime for an existing model, rather than a separate Qwen model release or a hosted subscription.
For someone evaluating local AI, the useful question is whether the whole computer can support the chosen model. GPU memory alone is not the budget: the installation documentation describes large system-memory allocations, model downloads, and optional disk caches. A gaming GPU does not remove those requirements.
Budget for the model, dependencies, and first start
- The installation guide recommends a current supported NVIDIA or AMD GPU with at least 12 GB VRAM. It says 8 GB NVIDIA configurations can run slowly; older hardware and Intel Arc have separate experimental routes.
- The CPU normally needs x86-64 with AVX2. Older CPUs without AVX2 are experimental. Standard installation targets Windows 10/11 and Linux, with automatic dependency setup documented for Ubuntu 22.04/24.04.
- Allow space beyond one model file: the guide lists roughly 70–80 GB for the model, about 6 GB for its draft layer, and another 1 GB when enabling images. Linux AMD setup may add about 10 GB for ROCm.
- On an AVX-512 CPU, Q2_0 can create an additional one-time expert cache of about 40 GB. An NVMe SSD is recommended, particularly for the first load and configurations that read model data from disk.
- First startup can make the computer sluggish while memory is allocated. Downloads can resume after interruption, and later launches reuse downloaded files. This still involves loading a large local model, not opening an instant cloud chat session.
Choose the version for your workload
The model guide separates compressed sizes from variants that change the model itself. Its RAM recommendations are a starting point, and available memory matters when browsers or other applications are open.
- For 32 GB RAM, the guide recommends Coder. It keeps half the experts, selected using code data, and is weaker outside coding and in languages other than English. With a 24 GB GPU, low-RAM mode can also accommodate selected full-model sizes.
- For 48 GB RAM, IQ2_XS or Q2_0 is recommended. With 64 GB, IQ2_XS remains the default recommendation, while IQ3_XXS and IQ3_S offer larger alternatives; IQ3_S leaves less room for other programs.
- The larger Unsloth variants can require substantially more memory and disk space. If data must be streamed from SSD during generation, a higher-quality size can be slower even though it technically fits.
The published model-size tables distinguish short-answer decoding, long-context decoding, and prompt processing. Those numbers answer different questions. The author's tested configurations and community reports are useful comparisons, but do not guarantee the same throughput on your PC. We have reviewed the documentation, not benchmarked Strata ourselves.
Start locally, then connect an app
Download the original repository and use its documented installer: START-HERE.bat on Windows or ./setup.sh on Linux. The local interface opens at http://127.0.0.1:8080. Source builds or particular backends can require additional tools, so check the installation guide for your exact hardware before assuming a prebuilt engine is available.
The README describes a browser Chat tab and a Monitor for model and hardware activity, plus compatible local API routes for external applications. The supplied product image is the author's Monitor screenshot beside a coding agent, not a Seailife test result. Compatibility with an API format should be checked against the specific client and runtime version you intend to use.
Check the platform limits before committing
Optional image input has backend-specific limits: the current README says AMD image input runs through the CPU on Linux and is not available on Windows yet. Requests queue by default; optional parallel processing trades capacity against per-answer speed. Local deployment is most relevant when you can manage the hardware and want applications to use a model running on your own machine. Keep the service local unless you have deliberately configured authenticated network access.