Skip to content

koboldcpp-1.125.1

Latest

Choose a tag to compare

@LostRuins LostRuins released this 22 Nov 07:54

koboldcpp-1.125.1

gtkobo
  • NEW: Added support from upstream for various new textgen models, including Inkling 2 Mini and Xing 4.0 29BA4B and more
    • Note: The Unsloth quants of inkling currently don't work correctly. Use the linked ggml-org quants instead.
    • On CUDA with partial offload and large prompts, the output for inkling2 can get corrupted. Known issue ggml-org#32710 and ggml-org#32778, use --no-op-offload to mitigate this.
  • NEW: Added support for upstream llama.cpp TTS models. For example, you can load Kokoro with the linked TTS model (put the mmproj in the --wavtokenizer slot). Some model families are duplicated across audiocpp, but all past implementations continue to work.
  • NEW: Added Qwen3ASR and SenseVoice transcription support
  • NEW: Added MMS and Qwen3 forced aligner models - when used together with a transcriber, it can generate timestamps for each spoken word. See this example use case.
  • TTS voice clone models that require a reference_text will now attempt to auto transcribe it if a transcriber model is loaded.
  • Fixed chatterbox TTS on cuda
  • NEW: Added integrated support for hosting a BeaverNet worker. BeaverNet is a free volunteer crowdsourced LLM provider run by @drummerv, and works similar to AI Horde. Anyone can run a worker for either platform via KoboldCpp to share your GPU and allow others to generate text with a model you are hosting.
  • Renamed --whispermodel to --transcribemodel
  • Reduced VRAM usage of YuE2
  • Bridged KoboldCpp native WebSearch into MCP - when websearch is enabled it will be available as a MCP tool.
  • Added support for gzip compressed HTTP requests
  • Standardized all request timeout values to read from --reqtimeout.
  • Added support for MiniMax H3 Image (get .kcppt template here), thanks @wbruna
  • Added support for interleaved reference image prompt support on some models, also added various image gen fixes by @wbruna
    • For example, if MOSS image was loaded with an appropriate Qwen 9B mmproj in --sdclip2, then you can add a reference image and refer to it in the prompt e.g. a book showing a portrait of <image:1> on the cover. This allows a basic form of editing on some models that don't normally support it.
  • Allow seamless resuming from an in-progress image generation if you disconnect. This requires Poor connection mode to be enabled. Also allow explicit aborts in poor connection mode.
  • Added basic support for the sd.cpp native sdcpp API - this works as a compatibility layer for tools written for that API
  • Added a dedicated delete tool for koboldcpp agent. In auto mode, it can be scoped within the workdir to run automatically without resorting to shell.
  • Also added a /attach command to attach images to the next message from disk in koboldcpp agent.
  • Make all Windows builds compatible with Windows 7
  • Added CI for experimental Windows ARM builds by @Harishmcw
  • Added a basic make install target
  • Prevent a potential XSS in noscript mode, thanks @1neZe0
  • Fixed some build issues by @askmyteapot
  • Fixed FreeBSD build
  • Fixed KoboldCpp Agent not being bundled in macOS builds
  • KoboldCpp Agent web requests follow the koboldcpp --nocertify mode if set.
  • Minor breaking change - Audio support has been dropped from LTX video generation. It didn't work very well and upstream did not merge the PR. Now that we have minimax h3 video, keeping this divergent feature is unnecessary.
  • Breaking Change: When --password is used now, all generation endpoints are secured by the API key. This means that things like noscript will not work when --password is set, preventing the loophole of using some functionality that did not support authentication. API key support has also been added to SDUI and KoboldAI Lite for image generation and interrogation via the A1111 endpoint.
  • Fixed a termux build issue
  • Increased total number and size per slot of server --savedatafile saves. Make server side saves atomic (thanks @Uberhumus).
  • Fixed an issue with unwanted prompt_overflow rejection for doable requests for AI Horde (thanks @tazlin)
  • Updated Kobold Lite, multiple fixes and improvements
    • Display used koboldcpp context token budget (thanks @Uberhumus)
    • Reworked save slots, now support up to 20 browser saves
    • Custom OpenAI endpoints can now optionally be used without a key
    • Allow choosing save location in corpo mode with a dropdown
  • Merged fixes, new model support, and improvements from upstream

Hotfix 1.125.1 - BeaverNet agent sets provider active on connect, added optional whole word WI keys

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 or check our koboldcpp.net website.