Strata released a hotfix, v0.1.43.1, to address speed reductions and crashes experienced by users of its 125-billion-parameter AI model. The update reverts the adaptive overlap setting to an opt-in basis after it caused significant slowdowns on certain NVIDIA graphics cards, and it fixes issues related to Linux NVIDIA decode performance that had dropped token rates significantly. Additionally, the hotfix resolves long prompt hangs on Windows, restores throughput for dual-card setups, and addresses various crashes across multiple hardware configurations. Overall, Strata aims to maintain output consistency while improving performance and stability for users running the model on consumer-grade hardware

Strata v0.1.43.1 Hotfix Fixes 125B Model Speed and Crashes

Strata, the open-source engine that runs a 125-billion-parameter AI model on a gaming PC, shipped v0.1.43.1 today as a hotfix for the previous day's release. The update reverts adaptive overlap to opt-in after it slashed decode speed by up to 21 percent on RTX 50-series desktops, and repairs a Linux NVIDIA decode collapse that dropped an RTX 4090 to single-digit token rates. It also restores roughly 35 percent throughput on two-card batch serving, fixes long-prompt hangs on Windows and SATA SSDs, and repairs regressions on AMD RX 6000 and Intel two-card setups. All changes keep answers byte-identical to v0.1.43, with one exception: a message quoting </tool_call> is now treated as plain text.

Strata v0.1.43.1 Hotfix Fixes 125B Model Speed and Crashes @ Linux Compatible