llama.cpp and its underlying ggml tensor library reached version 1.0 on November 13. The important part is not the round number. It is the project’s new rule that a breaking change to its public C API should require a new major version.[1]
That gives developers of local-AI apps, language bindings and Linux packages a useful signal they did not have during the 0.x era. It does not freeze llama.cpp, guarantee that every command-line option will survive, or make this particular upgrade effortless. In fact, the 1.0 release removes a pile of deprecated interfaces on its way through the door.[2]
What the promise covers
The published scheme is conventional semantic versioning: a major release can break the public C API in the project’s include directory; a minor release can add backward-compatible features, backends, model support or API calls; and a patch release is for bug fixes without an API change.[1] Under the SemVer specification, 1.0.0 is the point where a project defines its public API, while 0.x versions are considered development versions that may change at any time.[5]
For someone who only downloads a desktop app and chats with a model, this sounds like plumbing because it is. For apps built on the public C API, good plumbing makes a routine runtime update less likely to demand a surprise integration rewrite.
The benefit is more direct for maintainers. An app using the public C API can pin a supported major version instead of treating every upstream update as potentially incompatible. Package managers can express dependencies more sensibly. Debian packager Christian Kastner wrote in ggml-org’s versioning discussion that the change enables shared-object versioning and makes it practical to ship libllama in public library paths, where a future break can be communicated with a version bump.[4]
There are boundaries. The project’s precise table names the public C API, not every CLI flag, server endpoint, build option or internal header.[1] Apps that use private internals still own that risk. So do integrations that assume behavior the public interface never promised.
Version 1.0 can still break an old integration
A compatibility policy begins after the line is drawn. llama.cpp 1.0 removes llama_set_warmup, 35 deprecated functions, defrag_thold and the corresponding --defrag-thold option. It also moves the NextN speculative-decoding interface into the public llama.h header and renames one context-state function.[2]
The ggml release performs its own cleanup, deleting deprecated API and old DIAG_MASK operations. Its RPC protocol moves to version 9 so incompatible peers are rejected rather than quietly used together.[3]
That makes 1.0 a sensible target for a fresh integration, but not an automatic update for a working server. If you build llama.cpp or link libllama yourself, read the API-change list and compile your application in a test environment first. If a desktop app bundles the runtime, wait for that app’s maintainer to qualify the release. Replacing a known-good bundled library by hand defeats most of the stability this policy is meant to provide.
Packages are not all on the same track
Availability is already uneven. Homebrew listed llama.cpp 1.0.0 as its stable formula when checked on November 15, with ggml 1.0.0 as a dependency and binary bottles for Apple Silicon Macs plus ARM64 and x86-64 Linux.[6] Debian unstable still listed llama.cpp 0.9.0, even though it already splits the project into libllama development and runtime packages, tools and services.[7]
The upstream distribution notes explain another wrinkle: official releases use vX.Y.Z, but nightly development builds keep commit-number versions written as b[NUM]. The same notes said Winget and the project’s app packaging were still publishing nightlies, with a switch to official releases left on the to-do list.[4]
Check the channel, not merely the project headline. A package called “latest” may be a newer nightly rather than the stable 1.0 line, and a distribution package may lag behind it.
No new GPU is required
Semantic versioning changes integration risk, not model memory. Version 1.0 does not require a new GPU, add VRAM, or make a model fit where it did not fit before. Your hardware needs still come from the model size and quantization, context length, selected backend and how much work is offloaded to the GPU.
The release itself is substantial. It adds model architectures, FP8 KV-cache support, a host-memory cache for mixture-of-experts models, backend optimizations, server hardening and a redesigned web interface.[2][3] Those changes may matter for particular workloads, but the release notes do not establish one universal speedup. Treat 1.0 as a compatibility milestone with a large feature bundle attached, not as a benchmark result.
For home-AI users, the right move is pleasantly boring. App users should let their app vendor handle the transition. Package users should verify whether they are installing a stable tag or a nightly. Developers should migrate off deprecated calls, pin the major version they support and keep testing updates.
The promise will earn its value over the next few releases. If minor updates preserve the public C API and the next real break arrives as 2.0, local-AI software becomes easier to package and less brittle. For a runtime sitting underneath so many apps, predictability is a feature.