GGML
A tensor computation library developed by Georgi Gerganov, optimized for CPU inference. Implemented in pure C with zero external dependencies, it is cross-platform, has a low memory footprint, and serves as the underlying technical foundation of llama.cpp.
Core Facts
The GGUF format was released by the llama.cpp community in August 2023 as a replacement for the GGML format.
The GGUF format was designed by Georgi Gerganov, the founder of the llama.cpp project.
GGUF（GPT-Generated Unified Format）是由llama.cpp项目定义的模型文件格式
Timeline (last 90 days)
GGUF replaced earlier legacy model file formats such as GGML/GGMLv3.
The core change in Ollama v0.40.7-rc1 is the commit 'mlx: deduplicate CUDA runtime payloads', which moves the CUDA libraries shared by the GGML and MLX backends into a sibling CUDA payload directory to eliminate duplicate packaging.
Ollama introduces the MLX backend on macOS as an optional or default path, so a single distribution package needs to accommodate both the GGML and MLX backends.
GGML is the low-level tensor computation library of the ggml/llama.cpp ecosystem, implemented in C/C++, cross-platform, supporting CPU and multiple GPU accelerations (CUDA, Metal, Vulkan, etc.), and is Ollama's default path for running models on most platforms.
Ollama uses precompiled GGML kernels that target generic x86/ARM configurations and do not distinguish the SIMD width, cache hierarchy and memory bandwidth characteristics of specific chips.
GGUF has gradually replaced the older GGML format and become the common standard for local inference tools such as ComfyUI, LM Studio, and Jan.
llama.cpp achieves low-level performance optimization through its self-developed GGML tensor library (later evolved into the GGUF format), with its core innovation being multi-level weight quantization technology.
GGUF was designed by llama.cpp project founder Georgi Gerganov and replaced the previous GGML format in 2023.
whisper.cpp is a C/C++ implementation based on the GGML tensor library that supports pure CPU inference and can achieve real-time transcription on Apple M-series chips.
All Facts (20)
The GGUF format was released by the llama.cpp community in August 2023 as a replacement for the GGML format.
90%VerifiedThe GGUF format was designed by Georgi Gerganov, the founder of the llama.cpp project.
90%VerifiedGGUF（GPT-Generated Unified Format）是由llama.cpp项目定义的模型文件格式
90%VerifiedGGUF packages model weights, quantization parameters, vocabulary, and hyperparameters into a single file and supports memory-mapped (mmap) loading.
80%Verifiedllama.cpp是由开发者Georgi Gerganov于2023年发起的开源项目，使用纯C/C++实现了LLaMA系列模型的推理
80%Verified4-bit quantization in GGUF format can compress model weights from 32-bit floating point to 4-bit integers, reducing model size by approximately 8 times.
80%VerifiedThe GGUF format was introduced in 2023 by llama.cpp founder Georgi Gerganov and has become the de facto standard for local deployment of open-source models.
70%VerifiedThe GGUF format was introduced by llama.cpp, succeeded GGML, and supports mixed-precision quantization.
70%VerifiedGGUF supports multiple quantization levels ranging from Q2_K (2-bit) to Q8_0 (8-bit), and Q4_K_M uses grouped quantization technology to become the community-recommended default quantization configuration.
65%Unverifiedwhisper.cpp is a C/C++ implementation based on the GGML tensor library that supports pure CPU inference and can achieve real-time transcription on Apple M-series chips.
60%UnverifiedGGML is a C-language tensor computation library developed by Georgi Gerganov and is the core computational backend of llama.cpp.
60%UnverifiedGGUF replaced earlier legacy model file formats such as GGML/GGMLv3.
50%UnverifiedOllama introduces the MLX backend on macOS as an optional or default path, so a single distribution package needs to accommodate both the GGML and MLX backends.
50%UnverifiedThe core change in Ollama v0.40.7-rc1 is the commit 'mlx: deduplicate CUDA runtime payloads', which moves the CUDA libraries shared by the GGML and MLX backends into a sibling CUDA payload directory to eliminate duplicate packaging.
50%UnverifiedGGML is the low-level tensor computation library of the ggml/llama.cpp ecosystem, implemented in C/C++, cross-platform, supporting CPU and multiple GPU accelerations (CUDA, Metal, Vulkan, etc.), and is Ollama's default path for running models on most platforms.
50%UnverifiedOllama uses precompiled GGML kernels that target generic x86/ARM configurations and do not distinguish the SIMD width, cache hierarchy and memory bandwidth characteristics of specific chips.
50%UnverifiedGGUF has gradually replaced the older GGML format and become the common standard for local inference tools such as ComfyUI, LM Studio, and Jan.
50%Unverifiedllama.cpp achieves low-level performance optimization through its self-developed GGML tensor library (later evolved into the GGUF format), with its core innovation being multi-level weight quantization technology.
50%UnverifiedGGUF was designed by llama.cpp project founder Georgi Gerganov and replaced the previous GGML format in 2023.
50%UnverifiedGGML是Gerganov自研的张量运算库，专为CPU推理优化，采用纯C实现，可在ARM、x86等架构上运行
50%