koboldcpp v1.121 Released: New Features & Bug Fixes

September 16, 2026September 18, 2026

At a Glance

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

Overview

LostRuins has released the latest version “v1.121" of “koboldcpp", a C++ tool that allows users to easily run GGUF models. Boasting 11,713 stars and licensed under AGPL-3.0, koboldcpp embeds the KoboldAI UI and operates as a single-file executable without requiring installation.

This version introduces numerous features that significantly enhance multimedia generation and UI convenience, such as media reference support for Minimax H3, a revamped MusicUI, and the addition of a runtime LoRA selector in Stable-UI (SDUI). It also includes tool-calling fixes and bug patches, making the local LLM environment even more comfortable.

Key Changes

Here are the main changes that affect users of local LLMs and related features:

Enhanced Multimedia Generation

• Minimax H3 Media Reference Support: Users can now generate Minimax H3 videos by attaching audio clips and multiple reference images.
• MusicUI Revamp: The music generation UI (MusicUI) has been revamped for more intuitive operation.
• Video LoRA Support: LoRAs can now be utilized in video generation.
• LTX Audio Fixes: Bugs related to LTX Audio conditioning have been fixed.

Web UI Improvements

• Runtime LoRA Selector Added to SDUI: At http://localhost:5001/sdui, image generation LoRAs can now be displayed, added, and removed on the GUI during runtime.
• Image Generation Recovery Feature: A “Poor connection mode" has been added to SDUI, allowing generated images to be recovered even if the connection drops or the browser is accidentally closed.
• Cloudflare Tunnel Stabilization: Keep-alives are now sent to prevent tunnels from disconnecting during long video generation sessions and similar tasks.
• Other SDUI Improvements: Added a custom fields submission box, increased the maximum video frame count to 240 (default) or 480 (when expanded), added elapsed time display, set the dynamic LoRA default strength to 1, and fixed image carousel selection.
• Kobold Lite Improvements: Fixed attached vision images being rendered in the correct turns, improved auto-scrolling, fixed <t2i> tag handling inside thinking blocks, and added mini-toggle buttons for TTS and speech recognition.

LLM Inference and API-Related Fixes

• xHigh Reasoning Effort Support: Added support for higher reasoning effort (xHigh).
• Tool Calling Fixes: Fixed tool calling support for Kimi models and Deepseek V4 Flash models.
• --ffncpu Flag Added: Works similarly to llama.cpp’s --n-cpu-ffn, serving as an alternative method to partially offload layers.
• Adaptive-P Sampler Optimization: Sampler processing has been optimized.
• Jinja Tool Temperature Adjustment: Default temperatures are now set lower in Jinja tool requests when payloads include tools.
• Streaming Bug Fixes: Fixed race conditions that occurred during streaming output.
• Grammar Memoization Fixes: Fixed bugs related to grammar memoization.
• Segmentation Fault Prevention: Fixed an issue where segmentation faults occurred when model loading failed.

Other Tool and Launcher Improvements

• .safetensor Metadata Analysis: The --analyze flag can now be used to analyze metadata not only for .gguf files but also for .safetensor files.
• GUI Launcher Improvements: Added a feature to hide cache slots when smart cache is off, improved accessibility to the help menu, and revamped the SaveDataFile selection UI.
• Hugging Face Downloader Improvements: Model downloader behavior has been enhanced.
• Docker Image Updates: Supported the KCPP_LORA environment variable for cloud providers, enabling direct downloading of comma-separated LoRAs.

Specifications

Here are the model and hardware-related updates in this version:

• Supported Models: Minimax H3 (video generation), Kimi models, Deepseek V4 Flash models (tool-calling fixes)
• CUDA Version Change for Old PC Binary (OldPC): CUDA 11.4 has been adopted for binaries targeting older PCs and older NVIDIA GPUs.
• Upstream Integration: The latest fixes, new model support, and improvements have been merged from upstream projects.

How to Get It

Please update according to your environment using the methods below.

Windows Environment

1. Download koboldcpp.exe from the Releases page.
2. Run the downloaded koboldcpp.exe. The GUI will appear; configure the necessary settings (such as GPU Layers), load a GGUF model, and launch it.

Note: If you are using an older CPU or older NVIDIA GPU, try the oldpc version (Cuda11 + AVX1) instead. If you are not using a GPU, the lightweight nocuda version is available.

Linux Environment

On modern Linux environments, you can run the following command to install and launch it directly:

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

After installation, run ./koboldcpp from the terminal to start it.

Android Environment (Termux)

In the Termux environment on Android devices, running the following quick setup script will automatically build and install it:

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

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.

Sources

• https://github.com/LostRuins/koboldcpp/releases/tag/v1.121