koboldcpp-1.122.1

koboldcpp-1.122.1

kobold_agent.mp4

PSA: KoboldCpp needs YOUR help to fight a fake phishing site (click here for info)!

• NEW: Added an integrated KoboldCpp Agent - KoboldCpp now ships with a bundled Agentic harness. This works like an extremely lightweight replacement for tools like Opencode, Codex or Claude Code. Comes with 9 built-in tools, and a tiny system prompt of only 2k tokens including all tools.
  • To use it, simply enable it from the Admin tab in the GUI launcher, or add --agent to your launch flags.
  • KoboldCpp Agent can also connect to third party backends, or any OpenAI Chat Completions compatible endpoint.
  • Add more tools by loading a mcp.json file, MCP tools will be shared to the agent. Note: MCP tools execute on the KoboldCpp server, while Agent tools execute on the agent client.
  • Comes with 3 approval modes for tool calling confirmation: on/auto/off. Exercise caution when approving tool calls.
  • To function effectively, KoboldCpp Agent requires at least 28k ctx and 8k gen amount, though larger values are recommended. Recommend to have at least 12GB VRAM for a good experience.
  • You can download a .kcppt template to get started with Qwen 3.6 35BA3B here, simply load and launch in the latest KoboldCpp.
  • Supports AGENTS.md, context compaction and many more features
  • Run /help in the Agent to get more information
• Added the ability to specify ubatchsize independently from batchsize, similar to llama.cpp. Pipeline parallel flag removed, it will auto enable when ubatch is smaller than batchsize
• Make sliders more compact, added ubatch slider
• Remove some smartcache spam when in quiet mode
• Increased default contextsize.
• Generation failures during streaming return an error object
• Revert P40 increase VRAM usage issue
• Sync upstream sd.cpp by @wbruna
• Limit max image preview size to handle pixel space diffusers
• Added --autoswapthreshold, now you can set the size limit before autoswap triggers to prevent swapping when small models are being utilized.
• Lowered SD VAE tiling threshold to 512
• Fixed some tool calling parser issues
• Refactor of koboldcpp.sh build/run script by @henk717
  • Note for package maintainers: This release includes a breaking change to koboldcpp.sh. To shorten compiles and make it easier to build from source, koboldcpp.sh now attempts to optimize for the target platform it is run on.
  • What this means: If you use koboldcpp.sh script in your build system, you must now include KCPP_PORTABLE=1 as the environment variable if you seek to produce builds that work on systems outside of your own.
• Allow generation of audio codes when using LLM for lyrics (2 step process), improved music UI
• Added keepalive whitespace padding for tool streaming and text sync requests, similar to image gen
• Smarter handling of max output tokens when close to max context
• Updated Kobold Lite, multiple fixes and improvements
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.122.1 - Removed some spamming terminal logs, allow --agent to be used with --cli in a single terminal session.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.121

koboldcpp-1.121

nobold_mtv.mp4

Music generated by @Dampfinchen, clips for music video created entirely in KoboldCpp

• NEW: Media reference support added for Minimax H3! - You will now be able to attach audio clips and multiple reference images to generate minimax H3 videos. The music video attached to this video was created entirely in KoboldCpp (in segments of 7s each)
• Fixed a race condition during streaming.
• Fixed LTX Audio conditioning bug
• Added support for video LoRAs (thanks @wbruna)
• Swapped to CUDA 11.4 for OldPC release binaries (tested and verified by @henk717)
• Fixed a bug in grammar memoization (thanks @Reithan)
• Fixed tool calling support for Kimi models and Deepseek V4 Flash models
• Handle invalid numeric API parameters (thanks @yhz5613813)
• Fixed WebSearch UTF-8 handling (thanks @Erquint)
• --analyze feature can now analyze .safetensor file metadata in addition to .gguf files.
• Hide cache slots in GUI launcher if smartcache is off
• Made help menu in GUI launcher more accessible
• Added a flag --ffncpu which behaves like llama.cpp --n-cpu-ffn, an alternative way to partial offload layers
• Optimized Adaptive-P sampler
• Prevent segfault on model load failure
• Reworked GUI launcher's SaveDataFile selection UI
• Add support for xhigh reasoning effort
• Improved huggingface model downloader
• Jinja tools requests with tools in the payload have a lower default temperature
• NEW: Reworked MusicUI (Music Generation UI) for a more intuitive user experience.
• Multiple Stable-UI (SDUI) GUI Improvements:
  • NEW: Runtime LoRA selector in included stable-ui (http://localhost:5001/sdui) - Thanks to @wbruna, you can now easily view, add and remove image generation LoRAs from your selected LoRA directory at runtime in SDUI with a nice GUI selector.
  • Fixed SDUI Image carousel selection
  • Image Generation Recovery - Enable "Poor connection mode" in SDUI to allow recovering generated images even if you disconnect or accidentally close your browser.
  • Improved cloudflare tunnel stability in SDUI - Sends a keepalive to help prevent tunnel collapse during long requests (e.g. video generation).
  • Added box for sending extra custom fields in SDUI, increased max video frames to 240 (default) or 480 (extended toggle), show elapsed time
  • Set SDUI dynamic LoRA default strength to 1.
  • Prevent viewing generation progress status if genkey is incorrect
• Updated Kobold Lite, multiple fixes and improvements
  • Attached vision images now render in correct turns.
  • Improved autoscrolling behavior
  • Fixed <t2i> tag handling in thinking blocks.
  • Added mini toggle buttons for enabling TTS and voice recognition if detected.
• Merged fixes, new model support, and improvements from upstream

In addition to the above our docker image now has KCPP_LORA support for use with cloud providers, use this variable along with direct links to the lora's you wish to download, they can be separated by a comma.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.120

koboldcpp-1.120

• Minor fix to assistant gen prefills being triggered incorrectly
• Fixed a bug where failsafe mode was incorrectly selected
• Added support for DirectIO model load mode (--usedirectio), also mlock and mmap can now be combined.
• New model support: Qwen3.8-Flash-Next and Ling-3.0-flash are fully supported. Note that there are bad quants floating out there, so make sure you use the right files.
• Image generation sync and update from @wbruna, various image generation fixes
• Updated Kobold Lite, multiple fixes and improvements
  • Added custom user-configurable javascript tools by @hu-yijie, compatible with standard tool calling.
• Merged fixes, new model support, and improvements from upstream

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.119

koboldcpp-1.119

pasta.mp4

• NEW: Added support for Video generation and I2V with Minimax H3.
  • Requires 4 files as described in this docs. For ease of use, you can download the MiniMax H3 quick setup .kcppt template here
• NEW: Added support for Muse Glimmer, jinja templates and tool calling also supported. Qwen 3.8 is also fully supported now.
• Added support for DSpark and Dflash speculative decoding
• Increased max images and audio attachments supported to 64
• Added support for Mistral reasoning budget by @JulienJBO
• Autoswap name matching fixes by @Anai-Guo
• Increase max runtime image LoRAs supported from 4 to 10.
• Added support for Image progress and previews during image generation by @wbruna
• Support Autoswap for openai image endpoint, added support for OpenAI Image editing API /v1/images/edits
• Fixed an image gen memory leak
• Use python's CSV writer for benchmark saving
• Updated Kobold Lite, multiple fixes and improvements
• Merged fixes, new model support, and improvements from upstream

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.118.1

koboldcpp-1.118.1

• Row Split has been removed, selecting it will now default to tensor split.
• Rocwmma has been removed, matching upstream behavior.
• Added support for qwen3tts aliases for all voices.
• Breaking change for RPC to match upstream behavior, allowing llama.cpp RPC server and clients to be interchangeable with koboldcpp ones again.
• Fixed loading for SeFI image.
• Fixed a quality regression in Qwen Image and Qwen Image Edit.
• Fixed an issue with incorrect GPU selection when multi gpu is used.
• Fix for multi-turn deepseek v4 flash prompt process issues.
• Increased default server thread counts
• Added more information when running in batched (parallel requests) mode
• --jinjathink flag automatically enables jinja
• Added support for loading int8 convrot image models, exclude z-image from models that support image references.
• Image generation abort by @wbruna - Now when generating an image, you can abort the generation by closing the connection. This behavior can be disabled with the keep_image_gen_on_disconnect request field. Pressing cancel in SDUI will also do this.
• Also synced new version of sd.cpp from @wbruna
• Fixed generate more button being broken when whisper model is no longer loaded by @Anai-Guo
• Set no_host by default, please report if performance issues occur.
• Added multiple language support for qwen3tts, usable only with XTTS endpoint currently.
• Updated Kobold Lite, multiple fixes and improvements
  • Placeholder context handling improvements by @hu-yijie, this is not heavily tested so please report any issues faced.
• Merged fixes, new model support, and improvements from upstream

1.118.1 Hotfix - Set no_host to default false again.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.117.1

koboldcpp-1.117.1

• Fixed terminal output sometimes not showing thinking traces
• Multimodal Images are now inserted inline instead of at the top of the context. This should improve vision coherence and quality on most models.
• llama.cpp UI mic input is now disabled by default (toggle in settings)
• Prevent MTP drafting for parallel requests as it's not currently supported.
• Added Ollama compatible embeddings endpoint
• Added Ollama compatible streaming and tool calling. This make KoboldCpp fully compatible with all Ollama using software, simply set the port to 11434 and it should work out of the box.
• Merged stable-diffusion.cpp image gen updates, thanks @wbruna
• Save and display Image LoRA information when generating an image with pre-loaded LoRA.
• Allow viewing the last generated image if SDUI crashes or loses connection (requires the correct genkey)
• Updated Kobold Lite, multiple fixes and improvements
  • NEW: Group chat rework: Multi-character avatar and memory management has been added, thanks to @hu-yijie
  • Minor audio embed fix by @mosujiba
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.117.1 - Actually increase maximum defaultgenamount to 32k

Notice: This will be the last version that supports --splitmode row in CUDA as it has been removed upstream. For those use cases, you'll have to migrate to tensor split or layer split.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.116.1

koboldcpp-1.116.1

• NEW: Added Krea 2 Turbo support (Recommended). This is an extremely powerful image generation model with incredible knowledge.
  • Grab and load the Krea 2 Turbo template containing all necessary files here.
  • Recommended settings for Krea 2 Turbo: 8 steps at 1.0 guidance cfg. Prompt with descriptive natural language.
• NEW: Added Ideogram 4 support.
  • Grab and load the Ideogram 4 template containing all necessary files here.
  • Note: Ideogram4 prompts should be in JSON format. Read how to prompt it here.
  • For newbies, Krea2 is recommended instead. Thanks @wbruna for the sync and testing.
• NEW: Added Boogu Edit support. A simple but good image editing model.
  • Grab and load the Boogu Edit template containing all necessary files here.
• For more details on the image gen models and alternative sources, refer to these documents
• NEW: The built-in llama.cpp Web UI has been updated and now supports MCP servers, thanks to @henk717
• Changed to attempt to use llama.cpp jinja tool call response parser first. This will allow better out-of-box support for new tool calling models. Please report any issues encountered.
• Fix: MTP support has been fixed, and speeds should now be good across the board for both Qwen and Gemma assistant MTP.
• Added a failsafe target for MacOS builds that does not use Metal, only Accelerate. This can be enabled with --failsafe for MacOS builds.
• Fixed a regression with huggingface xet file download speed, this was needed due to changes on the HF side.
• Added an option to save TTS outputs as MP3. This can be done from the API as well, by setting response_format to mp3 as per OpenAI docs.
• Added an advanced CLI-only flag --allow-config-onready, which re-enables --onready commands from config files. This flag can only be manually triggered directly from the terminal and itself cannot be saved. This allows power users to run --onready commands when swapping .kcpps configs via API or the Web UI.
• Fixed drafting during save and reload states, allow drafting when mmproj is loaded (but avoid using them together). Default draft amount set to 4.
• Increase image gen prompt length limit to 3000 characters.
• Cap n_outputs for MTP (thanks @Pento95)
• Fixed a regression with Qwen VL mmproj vision quality.
• Load image generation weights eagerly
• Added Jinja toggle to the Quick Tab
• Added experimental support for uploading reference audio for LTX2.3 (Audio-To-Video)
• Updated Kobold Lite, multiple fixes and improvements
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.116.1 - Fixed an SSL issue with some xet downloads, fixed reasoning content not appearing when streaming was disabled.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.115.2

koboldcpp-1.115.2

sad.mp4

(Base Image = Klein 9B, To Video = LTX2.3, Add Music = AceStep 1.5, all 100% done in Kobold)

• NEW: Full support for the Anthropic /v1/messages API has been implemented, including multimodal and tool calling support.
• Added an easy toggle for default reasoning effort in the GUI launcher, can also be set with --reasoningeffort. This uses --gendefaults under the hood to set API defaults.
• Added support for the Gemma4 UV (12B) models and their vision/audio mmprojs.
• Improvements for smartcache implementation for RNN models.
• MTP and Gemma assistant models are now supported.
  • To enable MTP if it's built into the model, use --usemtp or the GUI launcher toggle. Otherwise, simply selecting the MTP drafter (e.g. gemma-4-26b-A4B-it-assistant-Q4_0.gguf as a --draftmodel along with the main model e.g. gemma-4-26B-A4B-it-UD-Q4_K_M.gguf is sufficient.
  • To see the drafting hitrate, use --debugmode. A good sample prompt is "Give me the first 100 integers".
• Increased max prompt length for image gen. Increased launcher context size defaults to 16k.
• Updates and various image gen fixes from @wbruna
• Accept more SSL file extensions
• Video generation reference image system overhauled.
  • You can now select videos to end on a specific frame in SDUI Img2Img, as well as specify both a start and end frame by uploaded 2 reference images in txt2img.
  • The reference image selector in SDUI has been overhauled and should work much better now.
  • Lower default VAE tiling threshold to 640, but you might want to lower it even more on low VRAM systems
  • Allow selecting up to 32 FPS if extended limits selected in settings.
• NEW: Allow limiting VRAM for image and video generation with --sdvramlimit. This will guarantee that you can load any image gen model by splitting the compute graph.
• If music gen was used without a --musicllm, fall back to a loaded text LLM.
• Reordered CLI args to be more alphabetical
• Retire the clip_quantize tool, since it can now be done with the regular gguf_quantize
• Updated Kobold Lite, multiple fixes and improvements
  • Fixed a bug with incorrect multimodal images/audio order sent to backend
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.115.1 - Sampling speedups, should fix any MTP slowdown issues. Overall sampling is also significantly faster at high tokens per second. Update is recommended

Hotfix 1.115.2 - Support new templates for LFM2.5, Cohere2Moe and adjustments to drafting.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.114.1

koboldcpp-1.114.1

kobo.mp4

• NEW: Experimental parallel text generation requests (continuous batching) is now optionally supported - Normal text gen requests can be executed in parallel instead of queued in this experimental mode. To use, set --parallelrequests X where X is the number of simultaneous generations to allow at one time, or set parallel requests in network tab in the GUI. This can be useful when hosting for many users or on Horde. Note that this is a very experimental feature, support is limited. (Thanks @AlpinDale)
  • Limitations: Only supports text gen requests with basic samplers, no multimodal, no special stateful samplers (e.g. antislop, grammar).
  • Limitations: Cannot be used together with context shifting or fast forwarding.
• NEW: RPC backend support is added - RPC allows multiple GPUs to be shared over the network, allowing for distributed generation with multiple computers! Note that it's not advised to expose RPC to the public internet! Set it up from the Network tab, or use --rpcmode and --rpctargets (if connecting). Default --rpcport is 5551. Thanks @Neresco for initial reference implementation.
  • --rpcmode connect = access a remote GPU
  • --rpcmode host = share your GPU
• NEW: Support for LTX2.3 video generation added! As usual, credits to @leejet from stable-diffusion.cpp for the original implementation. It works, although it's still somewhat buggy with some setups.
  • Quick Start Guide (16GB VRAM needed):
  • First, download the latest KoboldCpp.
  • Download and load this ltx2.3.kcppt template, open koboldcpp, load the config and select your preferred GPU backend (I used Vulkan). This will download 5 other models (ltx2.3-22b-distill, gemma3-12b, embedding-connector, audio-vae and video-vae), we will use the Distilled versions. Launch Kobold and select your backend.
  • Once loaded, launch http://localhost:5001/sdui when done (if you're remote, you can enable remote tunnel for a cloudflare link)
  • There are 2 options, txt2vid and img2vid. I'd recommend starting with a simple txt2vid, just setup a basic prompt, make sure steps are about 15 and guidance cfg is 1. Do not exceed 512x512 for now. 50 frames should work for a quick demo, that's about 4 seconds of video.
  • In SDUI, enable AVI downloads in settings.
  • Generate! It should take less than 1 min on a good GPU. When done, a GIF will be shown with the generated video.
  • To get the AVI version (with audio), click on the video, and select [Download AVI]
• Increased max frame limit, added FPS controls, added option to continue a generated video from final frame.
• NEW: Support for many other new Image Generation models are also added! (Thanks @wbruna)
  • Microsoft Lens image generation model support
  • HiDream o1 image generation model support
  • LongCat image generation model support
  • Ernie image generation model support
• Also added support for setting individual devices for CLIP and VAE for image generation (Thanks @wbruna)
• SDUI send as reference image toggle for img2img is now enabled by default.
• Allow loading custom TAE (Tiny Auto Encoder) image VAEs in the VAE file select.
• NEW: Multithreaded MP3 generation for AceStep - Multithreaded MP3s generate much faster, default also changed to 192kbps audio (thanks @askmyteapot)
• QoL Fix: You can now easily toggle thinking with --jinjathink [default/true/false] or using the GUI dropdown. Internally it just sets the thinking_enabled jinja kwarg.
• Jinja kwargs object can now be passed via API as chat_template_kwargs.
• Breaking change - make GPU ID list in the GUI launcher 0-based index (0,1,2,3) instead of 1-based (1,2,3,4) index to match the CLI.
• Breaking change - SWA is now enabled by default on all models that support it (was off by default previously). You can disable it with --noswa, or by unchecking the SWA checkbox in the GUI launcher. Models without SWA are unaffected.
• Autofit now accounts properly for the exact memory taken by mmproj files
• Revert AceStep VAE chunk size back to 256 (see #2224)
• Added more options to GUI context size slider
• Fixed a potential vulnerability when --onready was used with admin mode, .kcpps swapping can no longer trigger onready commands.
• Updated Kobold Lite, multiple fixes and improvements
  • Allow setting Sampler and Scheduler for ComfyUI remote endpoints, and A1111 compatible endpoints in Lite.
  • Added more image gen resolution options
  • Added aria labels for classic UI separators
  • Fixed a bug that broke chat avatars in chat mode
  • Reworked and unified the Import Character from Website URL option. Added support for BotBooru.
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.114.1 - Fixed VAE tiling for video gen (recommended, enable with --sdtiledvae 1), fixed auto GPU selection

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.

koboldcpp-1.113.2

koboldcpp-1.113.2

intermission edition

• Breaking change: Split mode is now set independently, in order to accommodate tensor split mode (experimental). This more closely follows the syntax used upstream with --split-mode
  • Example Before: --usecuda rowsplit
  • Example Now: --usecuda --splitmode (layer/row/tensor)
• Allow a runtime image LoRA directory to be easily selected, will allow all image gen LoRAs inside to be loaded via the <lora:filename:weight> prompt syntax
• Fixed q5_1 kv type not using the GPU correctly in CUDA
• Handle some instances of duplicate think tags
• Speed up Ace Step VAE decoding by increasing chunk size. If it doesn't fit, remember to use --musiclowvram
• Allow customizing the multiuser queue limit via the GUI
• Various minor updates and fixes for sd.cpp, thanks to @wbruna
• Qwn3TTS speedups by switching back to original precision
• Switched wav output from ulaw back to pcm16.
• Avoid saving deprecated args in new kcpps files
• Added --reqtimeout, which allows you to specify the desired timeouts for some requests like router mode.
• Improved screen reader accessibility for the MusicUI
• Fixed handling when ipv4 not available
• Handle seed-oss think format
• Fixed some config loading issues
• Fixed a bug where mmap used much more memory that expected due to CPU repacking.
• Fixed a issue when prefilling a turn in jinja mode, thanks @Reithan
• Updated Kobold Lite, multiple fixes and improvements
• Merged fixes, new model support, and improvements from upstream

Hotfix 1.113.1 - Fixed a regression with the multiuser flag. Support loading MTP models (though the actual speculative decoding pass is currently not enabled).

Hotfix 1.113.2 - Fixed a regression with router mode and some segfault freezing issues.

Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users.
If you have an older CPU or older NVIDIA GPU and koboldcpp does not work, try oldpc version instead (Cuda11 + AVX1).
If you don't have an NVIDIA GPU, or do not need CUDA, you can use the nocuda version which is smaller.
If you're using AMD, we recommend trying the Vulkan option in the nocuda build first, for best support. Alternatively, you can download our rolling ROCm binary here if you use Linux.
If you're on a modern MacOS (M-Series) you can use the koboldcpp-mac-arm64 MacOS binary.
Click here for .gguf conversion and quantization tools
Newer rolling experimental builds can be found here, these are auto-updated and may be unstable.

Run it from the command line with the desired launch parameters (see --help), or manually select the model in the GUI.
and then once loaded, you can connect like this (or use the full koboldai client):
http://localhost:5001

For more information, be sure to run the program from command line with the --help flag. You can also refer to the readme and the wiki.