llama.cpp reaches v1.0.0 with breaking API changes
The open-source inference engine’s first major release brings breaking API changes, wider model support and a harder-to-abuse server.
The open-source llama.cpp project has reached v1.0.0, with ggml-org warning that the release includes breaking public API changes alongside a much broader set of local-AI capabilities.1
What changed
The release removes deprecated public API functions and promotes the NextN/speculative API into `llama.h`, the public header used by applications integrating the inference engine.1 It also adds configurable Mixture-of-Experts cache controls and an OpenAI-compatible `/v1/audio/speech` endpoint, extending the project beyond text-generation workflows.1
The underlying ggml component has also moved to v1.0.0. The release adds support for new model and multimodal architectures, while hardening `llama-server` against SSRF, CORS, DNS-rebinding and MCP-tool-cancellation risks.1 Those security changes are useful signals for anyone exposing the server, although the release itself does not establish how severe each issue was in real-world deployments.
Why it matters
llama.cpp is a portable C/C++ inference engine designed to run across consumer hardware and enterprise servers. The project’s accompanying announcement says it now supports more than 150 architecture variants, a scale that helps explain why API stability and compatibility matter to downstream tools.2 Existing integrations that rely on deprecated functions may need changes, while newer applications gain a wider set of model and server options.
The release is therefore more than a routine version-number bump. It marks a cleaner public API, expands the kinds of models and modalities the project can handle, and makes its built-in server more suitable for deployments that need an OpenAI-compatible interface. Practical performance gains and the compatibility impact for any particular application remain unverified until maintainers and users test their workloads.
Our opinion
A v1.0 label is meaningful here because llama.cpp sits underneath a large portion of the local-AI ecosystem, not because the number magically guarantees stability. The breaking API work is a reasonable trade if it gives maintainers a clearer foundation, but downstream projects should treat this as a migration release rather than updating blindly. The strongest part of the announcement is the combination of wider architecture support and concrete server hardening; the real test will be whether popular wrappers absorb the changes without leaving users stranded.