KoboldCpp 1.120 Adds a New Way to Load Models, and Support for the Newest Open Ones

KoboldCpp, one of the simplest ways to run an AI model on your own machine, shipped version 1.120 on Saturday. If you have never used it: it is a single downloadable file, no installer and no Python environment to wrestle with, that loads an open model from disk and gives you a chat window in your browser. This release is small but useful, and it is a good excuse to explain a couple of things that sound intimidating and are not.

The headline addition is a DirectIO model load mode, switched on with --usedirectio. Loading a model means copying several gigabytes from your drive into memory before anything can happen, and by default the operating system tries to be clever about it by caching those bytes along the way. DirectIO skips that middle layer and reads straight from the drive, which on many systems means a faster start and less memory wasted on a cache you do not need. Related: mlock and mmap can now be combined, two older options that control whether the model gets pinned in memory or mapped lazily from disk. The release also adds day one support for two brand new open models, Alibaba’s Qwen3.8-Flash-Next and Ling-3.0-flash, and the maintainer adds a warning worth repeating: there are bad quantisations of these floating around, so take the files from a source you trust. Rounding it out are custom user configurable JavaScript tools in the built in Kobold Lite interface, compatible with the standard tool calling format, plus image generation fixes and the usual merges from upstream llama.cpp.

What’s actually going on here: releases like this are the unglamorous half of local AI, and they matter more than the model announcements do. A new open model is only useful once something on your computer can actually load it, and the gap between “weights published on Hugging Face” and “runs on a normal laptop” is filled by projects like this one, llama.cpp underneath it, and the people making quantised versions. Quantisation, if that is a new word, is the trick of storing a model’s numbers at lower precision so it fits in less memory, at a small cost in quality. Getting day one support for a model released three days earlier is the sort of thing that quietly decides whether open weights are a real alternative or just a press release.

What this means for you: if you have been curious about running a model locally, this is a genuinely low commitment place to start. Download one file for your system, grab a quantised model, and you have a private assistant that works offline and costs nothing per message. It will not match a frontier model from a big lab, and you should expect it to be slower and less capable, but for drafting, summarising and asking things you would rather not send to anyone’s server, it is often enough. If you already use KoboldCpp, the practical reason to update is the two new models and, on machines with fast drives and tight memory, the DirectIO flag. Everyone else can happily ignore this one and wait for the next release.

Sources

• koboldcpp-1.120 release notes (GitHub)
• koboldcpp 1.120 (Freedom.Tech)
• Qwen3.8-Flash-Next (Hugging Face)