Back to Gap Map Infrastructure / Core ML frameworks & libraries

GGML

open source / Overall score: 3.8

GGML is a low-level C tensor library for machine learning with no third-party dependencies, zero runtime memory allocations, integer quantization support, automatic differentiation, and optimizers. It runs cross-platform and is the tensor backend behind llama.cpp and whisper.cpp. It is maintained by ggml.ai, the company founded by Georgi Gerganov and now part of Hugging Face; much active development now occurs in the dependent llama.cpp and whisper.cpp repos.

Openness

5 high confidence

5.0

license: MIT(OSI)
source: public
core features withheld: no

MIT license body confirmed (Copyright 2023-2026 The ggml authors); full source public. The repository is public and unarchived, and the README describes the whole library with no paid, enterprise or hosted tier beside it, so source is public and the core ungated.

https://cdn.jsdelivr.net/gh/ggml-org/ggml@master/LICENSE recorded 2026-09-29
LICENSE is the verbatim MIT License text: "Copyright (c) 2023-2026 The ggml authors" and "Permission is hereby granted, free of charge, to any person obtaining a copy".
https://cdn.jsdelivr.net/gh/ggml-org/ggml@master/README.md recorded 2026-09-29
README gives the build-from-source route ("git clone https://github.com/ggml-org/ggml"), describes a "Plain C/C++ implementation without any dependencies", and lists no paid, enterprise or hosted tier; core changes are directed to the public llama.cpp repository.
https://ggml.ai recorded 2026-09-29
Project site states "The library and related projects are freely available under the MIT license" and that "In the future we may choose to develop extensions that are licensed for commercial use"; no commercial extension is described as existing.
https://github.com/ggml-org/ggml/blob/master/LICENSE recorded 2026-08-13
LICENSE file is the verbatim MIT License text, Copyright (c) 2023-2026 The ggml authors

Adoption

3 medium confidence

3.0

ggml is vendored as source into llama.cpp and whisper.cpp rather than published as its own package, leaving GitHub stars as the only comparable adoption signal. A star is not a download, so the reading stays directional.

https://api.github.com/repos/ggml-org/ggml recorded 2026-08-13
stargazers_count = 15159 for ggml-org/ggml
https://github.com/ggml-org/ggml recorded 2026-08-13
GitHub repo page for ggml-org/ggml, 15.2k stars

Capability

5 high confidence

5.0

Foundational low-level runtime enabling efficient on-device LLM/ASR inference.

https://cdn.jsdelivr.net/gh/ggml-org/ggml@master/include/ggml-opt.h recorded 2026-09-29
The ggml-opt.h header exposes training support: "GGML_OPT_OPTIMIZER_TYPE_ADAMW" and "GGML_OPT_OPTIMIZER_TYPE_SGD" optimizers, with a "opt_period" gradient-accumulation setting.
https://cdn.jsdelivr.net/gh/ggml-org/ggml@master/README.md recorded 2026-09-29
README describes a "Tensor library for machine learning" with "Plain C/C++ implementation without any dependencies", "2- to 8-bit integer quantization, plus MXFP4 and NVFP4 microscaling formats" and "Broad backend support - CPU, GPU, NPU, and browser".
https://ggml.ai recorded 2026-09-29
Site describes ggml as "a tensor library for machine learning to enable large models and high performance on" commodity hardware, "used by" llama.cpp and whisper.cpp.
https://github.com/ggml-org/ggml recorded 2026-08-13
README still describes a dependency-free C tensor library with integer quantization, autodiff and optimizers, backing llama.cpp and whisper.cpp

Verified 2026-09-27