Industry News
Strata 0.2.10: With 8GB of video memory, it is now possible to use 128K of context space. As a result, the additional load is being transferred to the memory and hard drive.
The open-source local large-model tool Strata has been updated to version 0.2.10. The most notable change is that a system with 8GB of video memory can now run Qwen3.8 IQ3_S while still maintaining a context size of around 128K (131,072 tokens). Previously, attempting to use a context size larger than 64K on systems with such limited video memory often resulted in startup failures; in those cases, the context size had to be reduced back to around 48K.
The solution lies in dynamic scheduling. When there is not enough video memory, the KV cache size retained on the GPU is automatically reduced; if necessary, some of that data is moved to system memory or even the hard drive. Although this approach frees up video memory, the associated trade-offs are clearly stated in the official recommendations: a minimum of 32GB of system memory is recommended, 48GB is more appropriate, and 64GB is ideal for optimal performance.
The difference in performance depends on the hardware used. When the official version operates within the 8GB memory limit using DFlash2, it achieves speeds of between 63 and 73 tokens per second. However, when a genuine Intel Arc A750 graphics card with 8GB of RAM is used, the system can start, but the performance only ranges from 12.7 to 14.8 tokens per second. The former scenario represents a memory limitation imposed on high-end cards, while the latter case more accurately reflects the performance of a device using 8GB of RAM.
Another significant contribution to saving system memory lies in DFlash2, which has been set as the default speculative decoder in this version. Previously, the MTP layer occupied approximately 5.2GB of system memory; however, DFlash2 requires only around 625MiB. Nevertheless, the developers have positioned DFlash2 as a solution that prioritizes single requests while still offering advantages in terms of multi-concurrency and hardware compatibility compared to MTP.
Users with multiple graphics cards receive the All-Cards Split Engine, which allows multiple cards to work together on processing each layer of data. When two RTX A6000 cards were used for code editing under Linux, the performance increased significantly, with the throughput rising from 195 token/s to 377 to 394 token/s.
An additional programming-specific tool, ISTA-DAS Coder, has been added. This tool was developed by the experts in video editing using DeepSeek-V4.1-Flash. However, it requires a significant amount of storage space due to its large weight and the size of the Engram data. Make sure your hard drive has enough free space before downloading it for the first time. Some minor bugs have also been fixed in the user interface; for example, when a response is too long, the approval button may get obscured, and you will be able to choose whether to continue generating the content or interrupt the process during its creation.