The open-source local AI tool Strata has officially released version 0.2.10. As a local large-model runner designed specifically for consumer-grade GPUs, this update brings breakthrough improvements in lowering VRAM requirements, increasing inference speed, and optimizing multi-GPU collaboration, making it easier for ordinary PCs to run hundred-billion-parameter large models.

Major VRAM savings: How can an 8GB GPU handle a 128K ultra-long context?

In previous local running scenarios, if users tried to enable a 64K or even larger context on an 8GB VRAM device, startup would often fail outright because the expert cache space was exhausted, and they sometimes had to compromise by dropping to around 48K.

In Strata 0.2.10, the official team optimized the low-level resource allocation mechanism: when VRAM is insufficient, the system prioritizes dynamically adjusting data allocation between VRAM and system memory, preserving the user-set context length as much as possible. This means users can now successfully launch large models such as Qwen3.8IQ3_S on an 8GB VRAM device and maintain an ultra-long context of up to 128K (131,072 tokens).

The core principle is that Strata automatically shrinks the KV cache window resident in the GPU based on real-time VRAM headroom, and when necessary, spreads part of the cache to system memory or disk. Official benchmark data shows that under an 8GB VRAM limit with DFlash2 mode enabled, generation speed can reach 63 to 73 tokens/s; even with a physical Intel Arc A750 8GB GPU, it can successfully launch a 128K context and maintain a generation speed of 12.7 to 14.8 tokens/s.

However, developers should also note: although pressure on VRAM hardware is shared by system memory and SSDs, this also places higher demands on the computer's memory configuration. The official recommendation is to start with at least 32GB of RAM, 48GB is more suitable, and 64GB is what truly allows you to handle various complex scenarios with ease.

DFlash2 accelerator enabled by default: saves an extra 4GB of VRAM

Another core highlight of this update is that the new DFlash2 has officially become the default speculative decoder.

As an efficient speculative decoding technology, the basic logic of DFlash2 is to have a lightweight auxiliary model predict upcoming generated content in advance, then hand it to the main model for batch verification. Compared with the previously used MTP speculative layer, which required about 5.2GB of VRAM, DFlash2's VRAM usage is greatly reduced to about 625MiB, saving about an additional 4GB of valuable VRAM from this one change alone.

In terms of generation speed, whether on Windows devices equipped with Ryzen AI Max, RTX PRO5000, or Apple M5Max devices, DFlash2 shows a clear single-request performance advantage. However, traditional MTP still has unique strengths in multi-concurrent processing and hardware compatibility, so the official team positions DFlash2 as the preferred strategy for single-request scenarios.

Multi-GPU workload splitting upgraded again: code editing generation speed nearly doubles

For hardcore enthusiast users with multiple GPUs, Strata 0.2.10 introduces the new All-cards Split Engine (multi-GPU joint split engine).

In the past, when running models on multiple GPUs, different layers of the model were usually mechanically assigned to different GPUs; the new version supports multiple GPUs processing every layer of the model together, allowing them to participate in parallel computation at the same time. Official tests show that when using two RTX A6000 GPUs to run a model under Linux, code editing task processing speed jumped directly from the previous 195 tokens/s to 377 to 394 tokens/s, nearly doubling performance.

Dedicated coding model and Agent experience improvements

To meet developers' advanced needs, the new version also introduces a coding-specific model called ISTA-DAS Coder. This model is based on DeepSeek-V4.1-Flash with expert pruning, retaining strong coding and Agent capabilities while reducing redundant overhead. However, because its model weights and Engram data files are relatively large, the first full download and run still require substantial storage space.

In addition, addressing interaction pain points in Agents and everyday graphical interfaces, version 0.2.10 includes comprehensive fixes and upgrades:

When an AI reply is long, the Approve and Deny approval buttons in the interface will no longer be obscured by page scrolling.

Sub-agents are allowed to receive follow-up instructions added by the user.

The automatic summary feature for chat history will prioritize matching the user's original language preference.

While AI is generating content, users can freely choose to wait, queue new messages, or interrupt directly.

Supports precisely quoting a previous conversation turn or selected specific text.

Improved the error prompt when Claude Code encounters context overflow, guiding it to try automatic compression.

Fixed a Windows bug where model loading became slow in certain background launch scenarios due to excessively low disk read priority.

Currently, Strata fully supports Windows, macOS, and Linux platforms. GUI users can upgrade directly to the latest version through the built-in update channel, while terminal users can complete the update through the relevant commands.