The Strata 0.2.10 release, launched on November 2, 2026, introduces the DFlash2 drafter as the default for the Qwen3.8-Flash-Next model, enhancing decode speed by 1% to 12% while freeing up VRAM. This release also features a new all-cards split engine that doubles throughput on dual NVIDIA RTX A6000 or RTX PRO 6000 setups and allows 8 GB cards to utilize a full 128K context without shrinking. The update has garnered significant attention, achieving approximately 38,700 GitHub stars, while the Strata engine aims to optimize memory usage by distributing model weights across GPUs rather than relying solely on VRAM. Despite the improvements, a high-end GPU, considerable RAM, and a robust CPU are still necessary for optimal performance, with the application available as a Linux AppImage, Windows installer, or Mac dmg
Strata 0.2.10 Release: DFlash2 Default, Multi-GPU Split Engine and 128K Context on 8 GB Cards
Strata 0.2.10, released November 2, 2026 by Niko Veit, makes the small DFlash2 drafter the default for the 125B Qwen3.8-Flash-Next model, freeing VRAM and adding roughly 1% to 12% decode speed depending on your hardware. Its new all-cards split engine, built from sFNX, launches automatically on Linux and roughly doubles throughput on dual RTX A6000 or RTX PRO 6000 Blackwell setups. For the first time, 8 GB cards get a full 128K context that shrinks nothing, while DeepSeek gains Metal-based vision support on Mac and a trimmed ISTA-DAS Coder variant appears. It's powerful on consumer hardware, though a smooth experience still wants a 5090-class GPU, 96 GB of RAM, and a 70 GB download.
