KoboldCpp
An open-source local large language model inference tool built on llama.cpp. It integrates image generation, speech recognition, and speech synthesis capabilities, and includes a built-in chat interface. The tool is known for its high degree of customizability, offering more than 56 adjustable parameters and 165 command-line flags. It is released under the AGPL-3.0 license.
Timeline (last 90 days)
Flash Attention is disabled by default in KoboldCpp, whereas in the llama.cpp source code it is enabled automatically by default.
KoboldCpp is a single executable file that can run AI models locally, generate images, process speech, recognize pictures, and comes with a built-in chat interface—zero installation, no persistent background service.
After downloading, KoboldCpp is a single file koboldcpp.exe of approximately 606 MB, while Ollama is an approximately 1000 MB installer that runs as a persistent background service.
On the release page, the Ollama installer has approximately 30 times the download count of KoboldCpp.
KoboldCpp's main developer has had 6 commits merged into the main llama.cpp repository, while Ollama's 4 core maintainers have collectively contributed 7 commits back to the engine.
KoboldCpp's single request structure has 56 tunable parameters, and the command-line launcher registers 165 parameters.
KoboldCpp treats the execution order of samplers itself as a configurable parameter, whereas Ollama's sampler chain is locked and cannot be modified.
In a dual-GPU setup with a 70-billion-parameter model, a user's real-world test showed llama.cpp reading prompt tokens at 1,490 tokens per second while KoboldCpp only achieved 34, the cause being a configuration error rather than an engine defect.
Both KoboldCpp and Ollama are built on the underlying llama.cpp engine.
KoboldCpp is labeled on its GitHub repository page as Forked from ggml.org/llama.cpp, released under the AGPL3 license with an MIT license to cover code inherited from llama.cpp.
6 more timeline events
All Facts (15)
Both KoboldCpp and Ollama are built on the underlying llama.cpp engine.
60%UnverifiedKoboldCpp's source code implements an Ollama-compatible endpoint that translates Ollama request formats and maps parameters, and can also respond to ComfyUI, transcription, and OpenAI endpoints.
60%UnverifiedAfter downloading, KoboldCpp is a single file koboldcpp.exe of approximately 606 MB, while Ollama is an approximately 1000 MB installer that runs as a persistent background service.
50%UnverifiedKoboldCpp's main developer has had 6 commits merged into the main llama.cpp repository, while Ollama's 4 core maintainers have collectively contributed 7 commits back to the engine.
50%UnverifiedKoboldCpp's single request structure has 56 tunable parameters, and the command-line launcher registers 165 parameters.
50%UnverifiedKoboldCpp treats the execution order of samplers itself as a configurable parameter, whereas Ollama's sampler chain is locked and cannot be modified.
50%UnverifiedIn a dual-GPU setup with a 70-billion-parameter model, a user's real-world test showed llama.cpp reading prompt tokens at 1,490 tokens per second while KoboldCpp only achieved 34, the cause being a configuration error rather than an engine defect.
50%UnverifiedFlash Attention is disabled by default in KoboldCpp, whereas in the llama.cpp source code it is enabled automatically by default.
50%UnverifiedKoboldCpp is a single executable file that can run AI models locally, generate images, process speech, recognize pictures, and comes with a built-in chat interface—zero installation, no persistent background service.
50%UnverifiedKoboldCpp is labeled on its GitHub repository page as Forked from ggml.org/llama.cpp, released under the AGPL3 license with an MIT license to cover code inherited from llama.cpp.
50%UnverifiedKoboldCpp has 56 adjustable parameters for a single generation request, and when launched from the terminal the launcher registers 165 different command-line flags.
50%UnverifiedKoboldCpp allows users to reorder the execution sequence of each position in the sampler chain, whereas Ollama's sampling chain is fixed.
50%UnverifiedIn a performance comparison case of running a 70-billion-parameter model with dual GPUs, the performance difference stemmed from the user mistakenly selecting the low VRAM option and disabling default settings such as flash attention, rather than the engine itself; performance was identical after adjustment.
50%UnverifiedKoboldCpp is a single 606 MB file that packages image generation, video generation, speech tools powered by Whisper, and a writing interface with persistent stories and character cards, spanning seven model families.
50%UnverifiedThe primary developer of KoboldCpp has 6 merged commits in the llama.cpp main repository; among Ollama's five core maintainers, four have contributed code to the engine, totaling 7 commits.
50%