KoboldCpp v1.123 Released: Major Agent Feature Upgrades

At a Glance
| Item | Value |
|---|---|
| Repository | LostRuins/koboldcpp |
| Version | v1.123 |
| Published | 2026-10-13 |
| License | AGPL-3.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
The latest version of KoboldCpp, v1.123, has been released. KoboldCpp is a C++ tool that allows users to run GGUF format AI models in a local environment and access them through a dedicated KoboldAI user interface (UI). It operates as a single, installation-free executable and is provided under the AGPL-3.0 license.
The most significant impact of this update for users is the substantial enhancement of the integrated “agent features." With the addition of granular confirmation modes and session save/restore functions, operating AI agents locally has become safer and more convenient. Users actively utilizing the agent features are advised to update promptly.
Breaking Changes & Deprecations
In this release, some startup options have been deprecated and integrated into new flags and configuration methods. If you are using previous settings, your startup script will need to be modified.
| Old Flag (Deprecated) | New Flag / Configuration Method | Details |
|---|---|---|
--jinjatools |
--jinja |
Using --jinja automatically enables Jinja tools and provides fallback handling upon failure. |
--multiuser |
--queuelimit |
Sets the limit for the connection waiting queue. Renamed to avoid confusion. |
These changes affect users who start the application from the command line with specific options. Please take note if you are setting up a multi-user environment or utilizing tool calling via Jinja templates.
Key Changes
Enhanced Integrated Agent Features and Improved Safety
Internal agent features in KoboldCpp have been significantly improved. The newly introduced “Confirmation Modes" allow you to control agent behavior using the /confirm command. For instance, setting it to auto mode automatically handles reading and writing files within the working directory while strictly requiring user approval for shell execution or operations outside the directory. This enables automation while ensuring safety.
Additionally, the /save and /load commands have been added for saving and restoring sessions. By specifying --agentload [filename] at startup, you can instantly resume from a previous session. A dedicated “Agent" tab has been newly added to the GUI launcher, making configuration easier.
Complete Overhaul of Batch Processing (Parallel Request) Mode
The batch mode handling parallel requests now fully supports standard samplers, grammars, tool calling, and vision functions. Users running multiple requests simultaneously can use the --parallelrequests N option to obtain multiple responses concurrently and without latency. No special reconversion is required to use this feature, but resources corresponding to the level of parallelism will be necessary.
Support for Qwen Image 2.1
The latest image generation and editing model, “Qwen Image 2.1," is now supported. This enables generating and editing images through KoboldCpp. To use this feature, please download and use the official dedicated template file (.kcppt). This adds a powerful option for users wanting to perform local image generation tasks.
Launch of Official Community Site
A new community site, “koboldcpp.net," has been launched. This site aims to provide resources for beginners starting with KoboldCpp and to prevent damage from phishing sites. Usage guides and contribution repositories are also available. Note that official executable downloads must still be done via the GitHub releases page.
Bug Fixes and Stability Improvements
An issue where out-of-memory (OOM) errors could occur due to repeat penalty range settings, as well as a problem where processing would halt during specific tool calls, have been fixed. Furthermore, an issue preventing the vision feature from working on older PC releases has been resolved. Since overall stability has improved, updating is recommended for all users.
Hardware Requirements
A new addition in this update is support for the image generation and editing model “Qwen Image 2.1." As mentioned earlier, using the dedicated .kcppt template enables image generation and editing directly on KoboldCpp. This has been incorporated as a result of upstream synchronization work.
Additionally, making samplers, grammars, tool calling, and vision features available in parallel (batch) mode means that features previously restricted to single requests can now be used in environments processing multiple requests concurrently. The documentation contains no new descriptions regarding initial support for specific hardware like GPUs or CPUs.
How to Get It
Because KoboldCpp is distributed as a single executable, you can update simply by downloading the latest version and replacing the old one. Get the appropriate file for your environment from the official Latest Releases Page.
- When using an NVIDIA GPU: koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux)
- If it fails to run on older CPUs or older NVIDIA GPUs: oldpc version (Cuda11 + AVX1)
- When not using an NVIDIA GPU: nocuda version
- When using an AMD GPU: The Vulkan option on the nocuda version is recommended. On Linux, rolling ROCm binaries are also available
- For Apple Silicon (M-series) Macs: koboldcpp-mac-arm64
After downloading, launch it by specifying startup parameters via the command line or by manually selecting a model in the GUI. Once started, you can connect to http://localhost:5001 to use it. Detailed startup options can be checked with the --help flag.
In a Termux (Android) environment, you can run the automatic installer with the following command:
curl -sSL https://raw.githubusercontent.com/LostRuins/koboldcpp/concedo/android_install.sh | sh
To build manually, follow these steps:
pkg update
pkg upgrade
pkg install openssl wget git python clang make
git clone https://github.com/LostRuins/koboldcpp.git
cd koboldcpp
make
python koboldcpp.py --model /path/to/model.gguf
If you wish to try experimental latest builds, you can obtain them from the Rolling Release, but please note that they update automatically and may be unstable.
Releases Since Our Last Article
Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.
| Version | Released | Release notes |
|---|---|---|
| v1.123.1 | 2026-10-13 | GitHub |
Related Articles
- KoboldCpp 1.124 Released: YuE2, TTS, and Image Gen Upgrades
- KoboldCpp v1.122 Released with Built-in Agentic Framework
What to Read Next
- Follow this tool → KoboldCpp overview and release history (137 releases tracked)
- Other inference engines and runtimes → llama.cpp / Ollama / vLLM

