KoboldCpp v1.122 Released with Built-in Agentic Framework

September 26, 2026September 27, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

KoboldCpp v1.122, the latest version of the AGPL-3.0 licensed inference engine for running GGUF models with ease, has been released. KoboldCpp operates as a single-binary application and is a C++-based inference environment that includes the browser-accessible KoboldAI UI.

The most significant change in this version is the integration of the lightweight agent foundation “KoboldCpp Agent". Designed with 9 built-in tools and keeping system prompt consumption down to around 2k tokens, it enables the creation of autonomous coding and task execution environments using local LLMs with minimal overhead.

Breaking Changes and Deprecations

This release introduces changes to the build script behavior specifications and streamlines several startup flags.

Maintainers distributing packages via koboldcpp.sh and users setting up build pipelines must explicitly set the environment variable KCPP_PORTABLE=1 to prevent generated binaries from being locked to the execution machine’s specific optimizations.

Key Changes

Addition of Built-in Agent Feature “KoboldCpp Agent"

A lightweight agent framework acting as an alternative to Opencode, Codex, Claude Code, and others has been bundled directly into KoboldCpp. Featuring 9 built-in tools, it keeps total system prompt consumption including all tool definitions down to approximately 2k tokens.

It can be enabled from the “Admin" tab in the GUI launcher or by specifying --agent in the startup options. It supports connecting to external OpenAI Chat Completions-compatible endpoints, and tool extension via the Model Context Protocol (MCP) is also possible by loading an mcp.json file (MCP tools execute on the KoboldCpp server side, while Agent tools execute on the agent client side). Three approval modes—on, auto, and off—are available for confirming tool calls. To run the agent practically, an environment with at least 28k+ context length, 8k+ generated token count, and 12GB+ VRAM is recommended.

Independent ubatchsize Setting and Pipeline Parallelism Automation

Similar to llama.cpp, ubatchsize (micro-batch size) can now be specified independently of batchsize. Accompanying this, a ubatch slider has been added to the GUI, and various slider UIs have been compactly rearranged. In addition, conventional flags for pipeline parallelism have been streamlined, changing to a behavior where they are automatically enabled when the ubatch value is set lower than batchsize.

Introduction of Model Swap Control Flag --autoswapthreshold

The --autoswapthreshold option has been added to specify the size limit that triggers automatic model swapping. This prevents unnecessary memory swaps from occurring when auxiliary smaller models are loaded and used.

Expansion of Default Context Size and Improved Context Management

The startup default context size has been increased, and the handling of maximum output generation tokens when approaching the maximum context length is now controlled more intelligently. Furthermore, changes have been made to return error objects if an error occurs during streaming generation, and keep-alive whitespace padding has been added for tool streaming and text synchronization requests.

Fixes for Image and Music Generation and Stable Diffusion Integration

Upstream sd.cpp synchronization has been incorporated, lowering the SD VAE tiling threshold to 512. Additionally, maximum size limits for image previews have been introduced to properly handle pixel space diffusion models, along with support for audio code generation (two-step processing) when using LLMs for lyrics generation, and improvements to music-related UIs.

Hardware-Specific Bug Fixes and Optimizations

The VRAM usage increase issue occurring in P40 environments has been reverted, and smartcache logging suppression during quiet mode operation, fixes to the tool call parser, and updates to Kobold Lite have been implemented.

Supported Models and Hardware

KoboldCpp v1.122 supports a wide variety of hardware environments and modern model architectures. By merging the latest improvements from upstream llama.cpp and sd.cpp, inference stability and supported ranges have been further enhanced.

Supported Models and Quantization Formats

This version enables inference for a broad range of LLMs (Large Language Models) provided in the GGUF format. Support for modern model architectures continues to be updated through upstream merges.

• Latest Model Support: To make maximum use of the new agent feature, .kcppt templates adapted for models such as Qwen 3.6 35BA3B have been released. This allows the latest high-performance models to be utilized as agents immediately.
• Image Generation Models: Through sd.cpp synchronization, image generation capabilities using Stable Diffusion have been enhanced. The VAE tiling threshold has been lowered to 512, and image preview size limits have been introduced to properly handle pixel space diffusion models.
• Quantization Tools: Tool suites for converting and quantizing into the GGUF format are also provided in connection with the project.

Supported Hardware Environments

Optimized binaries are provided to match users’ environments.

• NVIDIA GPU: Standard koboldcpp.exe (Windows) and koboldcpp-linux-x64 (Linux) utilizing CUDA 12 are recommended. Additionally, VRAM usage issues that occurred on specific hardware such as the Tesla P40 have been fixed, improving convenience for users of those GPUs.
• Older PCs and Legacy GPUs: An oldpc version (CUDA 11 + AVX1) is provided for older NVIDIA GPUs that cannot run the latest CUDA or older CPUs supporting only the AVX1 instruction set.
• AMD GPU: In environments without NVIDIA GPUs, using the Vulkan option with the nocuda version is recommended. For Linux environments, a ROCm-supported binary is also provided separately.
• Apple Silicon: For modern macOS with chips like M1/M2/M3, the koboldcpp-mac-arm64 binary is provided, enabling high-speed inference utilizing Metal acceleration.
• Android: Execution via the Termux environment is officially supported, enabling inference on mobile devices as well.

For package maintainers who build and distribute pre-built binaries themselves, the environment variable KCPP_PORTABLE=1 has been introduced to create environment-agnostic binaries.

How to Get It

KoboldCpp is designed around a “Zero Install" philosophy, meaning updates can generally be completed simply by downloading a new binary and replacing the existing one.

Usage on Windows

1. Download the latest koboldcpp.exe from the releases page.
2. Run the downloaded file. The GUI launcher will start; select your model, configure settings (such as GPU layer count), and click “Launch".
3. If you wish to configure detailed settings from the command line, you can check available options with koboldcpp.exe --help.

Usage on Linux

The recommended method is to download the koboldcpp-linux-x64 prebuilt binary from the releases page. After downloading, grant execution permissions (chmod +x) and run it.

To fetch and install the latest version directly from the terminal, the following command can be used:

curl -fLo koboldcpp https://github.com/LostRuins/koboldcpp/releases/latest/download/koboldcpp-linux-x64-oldpc && chmod +x koboldcpp

After installation, you can launch it with ./koboldcpp.

Usage on Android (Termux)

If you are using Termux on an Android device, running the following quick setup script will automatically handle everything from installing necessary dependencies to building:

curl -sSL https://raw.githubusercontent.com/LostRuins/koboldcpp/concedo/android_install.sh | sh

If you prefer to build manually, clone the repository and run the make command. To create a portable build shareable with other devices, you must build with the LLAMA_PORTABLE=1 flag.

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.

What to Read Next

• Follow this tool → KoboldCpp overview and release history (132 releases tracked)
• Other inference engines and runtimes → llama.cpp / Ollama / vLLM

Sources

• https://github.com/LostRuins/koboldcpp/releases/tag/v1.122
• https://github.com/LostRuins/koboldcpp/blob/concedo/README.md