The open-source local AI large model running tool Strata has received a major update with the 0.2.0 beta. As an open-source tool dedicated to letting ordinary home computers smoothly run large language models such as Qwen3.8-Flash-Next and Qwen3.6-35B-A3B, Strata not only introduces a brand-new standalone graphical interface in this upgrade, but also further and substantially lowers the hardware threshold for local deployment.
In terms of hardware configuration requirements, the new version deeply optimizes the combined usage of memory and VRAM (RAM + VRAM). A total of as little as 27.6GB of memory and VRAM is needed to run local large models smoothly. This means that if users have a graphics card with 12GB of VRAM paired with 20GB of system memory, the minimum deployment requirement can be met. In the specific recommended model tiers, the IQ2_XS recommended tier, which previously required a 48GB configuration, has dropped to 40GB, while the IQ3_XXS high-quality recommended tier, which previously required 64GB, has dropped significantly to 48GB, reducing overall hardware requirements by 8GB to 16GB.
Beyond the lowered hardware threshold, Strata 0.2.0 also achieves all-around hardware acceleration in inference performance. Among them:
NVIDIA camp: RTX 30 series graphics cards see text generation speeds improved by up to about 13% (measured: RTX 3060 improved 10.2%, RTX A4000 improved 13%);
AMD camp: RX 9000 series graphics cards see prompt processing speeds nearly doubled on some models;
Intel camp: the Arc A750 graphics card also achieves a significant improvement of about 42% in prompt processing speed.
In terms of interaction experience and ecosystem expansion, Strata 0.2.0 breaks free from the previous complexity and cumbersomeness of command-line compilation and deployment, fully supports Windows, macOS, and Linux platforms, and adds a native Chinese graphical interface with intelligent installation guidance. Even more noteworthy, the new graphical interface deeply integrates an AI coding Agent, supporting file reading and modification, web search, and automated task execution; meanwhile, image understanding is enabled by default for NVIDIA and AMD graphics cards, and the model download mechanism has been optimized with multi-threaded concurrency and resumable downloads.