llama.cpp reaches version 1.0 with a compatibility contract
ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026. The open-source project now sets clear rules for API and ABI breaks, with direct impact for integrators.
ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026, introducing a key feature for developers: major changes will now explicitly signal source and binary compatibility breaks. For integrators relying on this local inference stack, the message is clear: updates will no longer play hide-and-seek with dependencies. The project, used to run models on consumer machines as well as servers, thus shifts from a rapid evolution logic to a more readable contract.
Une version 1.0 qui dit ce qu’elle casse
The rule announced by maintainers is simple: a public C API break triggers a major version change, while a backward-compatible improvement is a minor version, and a fix without API change remains a patch. This distinction applies to libllama, the model runtime, and libggml, the tensor layer underlying computation. In other words, the version number no longer just counts releases but warns of surprises.
This formalization comes after 3.5 years of development and as the library has become a foundation for many local deployment tools. The project now claims support for over 150 model architecture variants and all major speculative decoding methods, an acceleration technique that proposes multiple tokens ahead to save inference time.
Ce qui change dans le moteur, le serveur et l’interface
Version 1.0.0 is not limited to numbering policy. On the API side, it removes 35 deprecated public functions and promotes the NextN interface from the llama-ext.h extension header to the public llama.h API. This promotion also brings speculative decoding under the compatibility rules.
Functionality-wise, the stack adds FP8 key-value caches for attention, a host memory cache for MoE models with configurable CPU share, and loading of external PLE or engram data from separate GGUF fragments. Supported models expand with Kimi-K3.1-Pro, Ripple, xing4_0, drama-3 for speech synthesis, and vision capabilities for MageVL-Cowork and Kimi-K3.1.
On the llama-server side, the team fixes server-side request forgery, CORS, and DNS rebinding vulnerabilities, while limiting HF_TOKEN to huggingface.co hosts. MCP tool calls are canceled if the client disconnects or expires, with notification to the MCP server. A new OpenAI-compatible route, /v1/audio/speech, completes the audio interface. The web interface was also revamped, with named discussion groups, a live-preview Markdown editor, mobile layouts adapted to safe screen areas, and full-screen dialogs. Yes, the Android keyboard finally cooperates too.
Pourquoi la compatibilité devient un sujet très concret
For package maintainers and application developers, this change matters as much as new functions. The promise of a major version signaling a break reduces ambiguity during updates, especially in environments where llama.cpp is installed via Debian, Homebrew, or other system channels. The process note dated August 7, 2026 adds a safeguard: an official version tag is created only if the internal copy of ggml matches the released version, to maintain compatibility with a system ggml installation.
The same movement is found in ggml v1.0.0. The project removes deprecated APIs and legacy DIAG_MASK operations, adds configurable YaRN truncation and FP8 support, and avoids sending scheduler inputs that are not read. CUDA and Metal backends gain quantized matrix multiplications on few rows and fusion work for MoE, while Vulkan on Intel, SYCL, OpenCL on Adreno, WebGPU, and Hexagon also receive additions. In short, hardware support continues to expand but with more discipline around interfaces.
Une base locale toujours plus large, mais moins permissive avec les intégrations
The project remains designed for local model execution, with portable C/C++ code, quantization, and optimized backends for machines ranging from individual workstations to enterprise servers. Meanwhile, pre-release artifacts cover Linux, macOS, Windows, Android, and iOS, while Docker, Debian, Homebrew, conda-forge, MacPorts, Winget, and XCFramework broaden distribution. That’s a lot of platforms to juggle, so the slightest compatibility error quickly becomes a headache in maintenance rooms.
To support this pace, the team says it relies on Hugging Face Buckets and Jobs to increase continuous integration capacity on master branches, with self-hosted runners provided by NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel. The project also indicates integrating LLMs into its own development cycle, while emphasizing three priorities: code quality, architecture control, and fine understanding of implementation details. For users, this translates into a more robust stack. For integrators, a rulebook finally written in black and white.
Key points
- llama.cpp v1.0.0 and ggml v1.0.0 were released on November 13, 2026.
- API and ABI breaks now require major version changes.
- The project claims support for over 150 model architectures.
- 35 deprecated functions removed from the public API.
- Artifacts cover Linux, macOS, Windows, Android, and iOS.
- ggml-org requires identical ggml copy for official tags.
By the numbers
- November 13, 2026 — release date of llama.cpp v1.0.0 and ggml v1.0.0.
- 35 functions — deprecated public API removed in llama.cpp v1.0.0.
- Over 150 variants — model architectures supported by the project.
- 3.5 years — development duration cited by maintainers.