github.com via Reddit

Koboldcpp v1.120 adds DirectIO load mode, Qwen3.8-Flash-Next

Open Source Inference ai-business

TL;DR

  • Koboldcpp v1.120, released August 29, adds a `--usedirectio` model load mode that can now be combined with mlock and mmap.
  • Full support lands for two new MoE models, Qwen3.8-Flash-Next and Ling-3.0-flash, with a warning about bad quantizations circulating.
  • Kobold Lite gains custom user-configurable JavaScript tools compatible with standard tool calling, plus image-generation fixes.

Koboldcpp shipped v1.120 on August 29 with a new DirectIO model load mode. The release notes describe it as "DirectIO model load mode (`--usedirectio`), also mlock and mmap can now be combined."

The build adds support for two recently released MoE models, Qwen3.8-Flash-Next and Ling-3.0-flash, with a cautionary line in the notes: "there are bad quants floating out there, so make sure you use the right files." The Ling-3.0 addition lands the same week Tencent open-sourced Hy4-preview, a 770B MoE with a 1M-token context, part of a steady run of large open MoE releases local runtimes have been racing to support.

On the UI side, Kobold Lite picks up "custom user-configurable javascript tools" compatible with standard tool calling, alongside an image-generation sync from contributor wbruna. Two smaller fixes land as well: assistant generation prefills that were being triggered incorrectly, and a failsafe mode selection bug.