Open sourceggml-orgConfirmed
Story 3 of 15 in the Friday 13 November editionggml-org releases llama.cpp v1.0.0 and ggml v1.0.0 with stable C APIs
Shipped on 13 November 2026 across Linux, macOS, Windows, Android and iOS, both C/C++ local inference libraries adopt strict semantic versioning across more than 150 model variants.
Edited by Ashutosh Shrivastava
14 Nov 2026, 09:50 IST · 2 min read
Georgi Gerganov and the ggml-org maintainers at Hugging Face released llama.cpp v1.0.0 (libllama) and ggml v1.0.0 (libggml) on 13 November 2026, marking the portable C and C++ local inference runtime’s first 1.0 milestone 3.5 years after its initial commit. Starting with v1.0.0, both repositories enforce strict semantic versioning across all public C headers in llama.cpp/include/ and ggml/include/, requiring a major-version bump whenever a change breaks API or ABI compatibility.
Semantic versioning contract, dynamic ggml backends and CI hardware matrix
Minor-version increments cover backward-compatible capabilities, quantisation formats, hardware backend optimisations, public header additions and new model architectures, while patch increments remain restricted to bug fixes with zero public interface changes. Pre-built release artefacts span Linux, macOS, Windows, Android and iOS, supporting more than 150 language and multimodal model variants alongside mainstream speculative decoding methods and dynamically loaded ggml backends discovered at runtime.
Continuous integration runs across Hugging Face Buckets, Hugging Face Jobs and self-hosted hardware runners maintained in partnership with NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.
Technical details
- Release versions and libraries
- llama.cpp v1.0.0 (libllama) and ggml v1.0.0 (libggml) under MIT open-source licensing
- Protected public interfaces
- Public C headers under llama.cpp/include/ and ggml/include/ governed by API and ABI SemVer
- Supported operating systems
- Linux, macOS, Windows, Android and iOS with runtime-loadable dynamic ggml compute backends
- Architecture and CI breadth
- 150+ model variants tested across NVIDIA, AMD, Qualcomm, Arm, IBM and Intel hardware runners
Release details & context
Shared-library upgrades for downstream Python, Rust, Swift and Go bindings
Prior to 1.0.0, rapid rolling-tag updates in llama.cpp frequently altered struct layouts or function signatures in llama.h, forcing downstream GUI launchers, game engines and FFI wrappers in Python, Rust, Go and Swift to pin exact commit hashes.
Locking ABI compatibility across 1.x minor and patch releases means applications linking dynamically against shared libllama and libggml binaries, combined with runtime-loaded GPU and NPU backend plugins, can adopt new model architectures and kernel speedups via drop-in library swaps without recompiling host code.
Common questions
Under llama.cpp v1.0.0 semantic versioning, do new model architectures require a major 2.0 release?
No, backward-compatible model support, quantisation schemes and public header additions ship in 1.x minor versions, while major bumps are reserved for breaking C API or ABI changes.
Can applications ship a lightweight llama.cpp binary and load hardware-specific ggml backends at runtime?
Yes, libggml supports dynamically loaded backend modules so distributors can load CUDA, Vulkan, Metal, SYCL or NPU backends at runtime based on the user’s device.
Sources
- PrimaryGitHub Release: ggml-org/llama.cpp v1.0.0github.com/ggml-org
- OfficialGitHub Release: ggml-org/ggml v1.0.0github.com/ggml-org
- OfficialHugging Face Blog: llama.cpp v1.0.0: A stronger foundation for the local AI ecosystemhuggingface.co/blog
- On XGeorgi Gerganov (@ggerganov) v1.0.0 announcement on Xx.com
Related open source launches
- Expo open-sources Exact to compile Contract UI specs to native OS views and a 20 KB web runtime
Expo open-sources Exact under the MIT licence, compiling declarative Contract UI specs to a 20 KB web runtime and native macOS, iOS and Linux views.
- MiniMax releases open-weight H3-Image for 10-second native 4K generation
MiniMax open-sources the 33.1B H3-Image model on Hugging Face, rendering native 16-megapixel 4K images in 6 steps and roughly 10 seconds.
- Runway open-sources 27.2B Praxis-1 world action model under Apache 2.0 for robot control
Runway open-sources Praxis-1 under Apache 2.0, topping open-source robot models on NVIDIA’s RoboLab-120 benchmark with 46.2% success.
- Thinking Machines Lab releases Inkling 2 Small weights on Apache 2.0
Thinking Machines Lab releases open weights for its 594B-total, 17B-active Inkling 2 Small multimodal MoE on Hugging Face under an Apache 2.0 licence.