Strata has released version 0.2.4 of its open-source local AI engine, enabling users to run the 763-billion-parameter DeepSeek V4.1-Flash model on a PC equipped with a single 12 GB graphics card and 46 to 64 GB of RAM. The engine efficiently distributes the workload across the GPU, CPU, RAM, and SSD, allowing for offline operation without telemetry. While the model is only accessible through the Terminal in this release, it comes in three compressed sizes ranging from 239 GB to 274 GB, with performance varying based on hardware capabilities, achieving approximately 8 to 10 tokens per second on mid-range setups. Strata's unique architecture allows it to scale significantly while maintaining a commitment to user privacy, with installation available through simple commands for various operating systems
Strata 0.2.4 Released: Run a 763-Billion-Parameter AI Model Locally on Your Gaming PC
Strata, the open-source local AI engine, released version 0.2.4 today, letting you run the 763-billion-parameter DeepSeek V4.1-Flash model on a PC with a single 12 GB graphics card and 46 to 64 GB of RAM. It does this by spreading the workload across your GPU, CPU, RAM and SSD, keeping the most-used parts in memory and streaming the rest from your drive. The model is Terminal-only in this release, comes in three heavily compressed sizes from 239 GB to 274 GB, and tops out at roughly 8 to 10 tokens per second on a mid-range RTX 3060. It's free under MIT with no telemetry, though output stays slow on 12 GB cards and the desktop app doesn't get DeepSeek until a later update.
