Strata 0.2.10 enables massive context windows on consumer hardware with vram optimization

The open-source local AI tool Strata has released version 0.2.10, introducing significant advancements for running large models on consumer GPUs by optimizing VRAM usage and inference speed. This update allows users to run models like Qwen3.8IQ3_S with an ultra-long context of up to 128K tokens on an 8GB VRAM device. This is achieved by dynamically shifting data allocation between VRAM and system memory when resources are constrained. Furthermore, the new DFlash2 speculative decoder is now the default, reducing VRAM consumption from approximately 5.2GB (used by the previous MTP layer) down to about 625MiB, saving an additional 4GB of VRAM. For multi-GPU setups, the All-cards Split Engine allows parallel computation across all GPUs for every model layer; testing with two RTX A6000 GPUs showed code editing speed nearly doubling from 195 tokens/s to 377-394 tokens/s. Beyond performance, the release includes a coding-specific model, ISTA-DAS Coder, based on DeepSeek-V4.1-Flash. System stability was also improved with fixes for interface issues, such as preventing approval buttons from being obscured during long AI replies, and better handling of context overflow errors in Claude Code. The tool remains compatible across Windows, macOS, and Linux.