llama.cpp 1.0.0 after 3.5 years: semantic versioning, 35 deprecated functions gone, FP8 KV cache. What the stability promise changes for local AI
The engine under Ollama, LM Studio and most on-device inference now promises that breaking changes arrive only with a major version; for the teams that link libllama, that is the end of a fortnightly maintenance tax.

On Thursday Georgi Gerganov's ggml-org released llama.cpp 1.0.0 and ggml 1.0.0, three and a half years after the project began as a weekend port of Meta's first Llama weights to a MacBook. The headline is not a feature. It is a promise: from now on, libllama and libggml follow semantic versioning. A major version means a breaking change to the public C API or ABI. A minor version adds backends, model architectures or new API without breaking old callers. A patch fixes bugs and changes nothing else. An official llama.cpp release is only tagged when its bundled ggml matches a released ggml, so system-wide installs from Debian or Homebrew stay consistent.
Why that matters more than it sounds
llama.cpp is the inference engine beneath a large share of local and on-device AI: Ollama, LM Studio, Jan and dozens of apps link it or shell out to it, and the GGUF format it introduced is how quantised open-weight models are distributed. Until now the project shipped a numbered build almost daily and reserved the right to change its API at any time. Teams that depended on it either froze a commit and fell behind on model support, or tracked the head and absorbed breakage. The release post acknowledges this directly, noting that many downstream applications depend on the work and that the maintainers will prioritise code quality and architectural control.
What is in 1.0.0
The deprecated public API is removed, 35 functions in all, so anything still calling them must migrate. The NextN speculative-decoding API moves from an extension header into the public llama.h. Memory handling gains FP8 KV caches and a host-memory cache for mixture-of-experts models with a configurable CPU share, which is the feature that lets a sparse model larger than video memory run by keeping cold experts in system RAM. New model support includes Kimi K3.1 Pro and vision for Kimi K3.1, a text-to-speech model and several text architectures, taking the supported count past 150 variants. The server fixes server-side request forgery, CORS and DNS-rebinding issues, restricts the HF_TOKEN to huggingface.co hosts, and adds an OpenAI-compatible speech endpoint. Pre-built packages cover Linux, macOS, Windows, Android and iOS through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework. Nvidia, AMD, Qualcomm, Arm, IBM and Intel supply self-hosted CI runners; Hugging Face infrastructure handles the rest.
The counter-argument
A version number is a developer commitment, not a code freeze. The release post is explicit that change continues; what changes is the signalling. End users of Ollama or LM Studio will notice nothing on Friday. And semantic versioning cuts both ways: a project that has moved fast for three years may find that the next genuinely necessary API break now requires a 2.0 and a migration guide, which is friction the daily-build era never had. Whether the maintainers hold the line through the next wave of architectures, this week alone brought a 594B NVFP4 model and a 1.6T MoE, is the test.
What to do
If you ship software that links libllama: move to the 1.0.0 tag, run your test suite against the removed deprecated functions, and set your dependency to 1.x with minor-version updates allowed. If you distribute GGUF models: nothing changes in the format, but check that your quantisation recipes still load. If you are evaluating on-device or on-premises inference for a product: the FP8 KV cache and MoE host cache are worth benchmarking against your current serving stack on the models you actually run, with attention to quality at long context, not only tokens per second. If you are an executive weighing self-hosting: the runtime's stability was the argument against; the GPU bill and the serving-stack gaps we described for Inkling 2 Small on Thursday are the arguments that remain.
The story so far
- Nov 14, 2026 IBM's Granite 4.1 LLMON previews train the model to separate instructions from data, claiming near-zero indirect prompt-…
- Nov 13, 2026 Inkling 2 Small's weights are out: 394 GB, Apache 2.0, 500 GB of VRAM to run, and no released serving stack supports it …
- Nov 11, 2026 Skild AI's Steel runs robot simulation up to 13x faster than PhysX on mixed batches and uses 62% less memory; here is wh…
- Nov 10, 2026 Nvidia's GR00T 2 claims 71% one-shot success on 50 tasks. Here is what that number does and does not say
Key terms in this story
- Semantic versioning A version-number convention (MAJOR.MINOR.PATCH) in which breaking changes bump the first number, compatible additions the second and bug fixes the third.
- Open-weight model A model whose trained parameters are published for anyone to download and run.
- Inference Running a trained model to produce output. What you pay for when you call an API.
- Token The unit models read and write; roughly three-quarters of an English word. Pricing and context limits are counted in tokens.
- NVFP4 Nvidia's 4-bit floating-point weight format, used to train and serve very large models on Blackwell-class GPUs with roughly a quarter of the memory of 16-bit weights.
More in Developer tools
All →
Perplexity's Brain: an agent memory that records its own work and rewrites its skills overnight, with correctness up 25% on repeat tasks. What to copy, and what to guard
Memory about the user is old news; memory about what the agent did, what failed and what was corrected is the new layer, and it is also a new place fo…

MAI Code 2 lands in GitHub Copilot: Microsoft says 98% of GPT-6.1 Sol at 40% of the cost and 96% of Claude Opus 5.5 at 5%. What the claims hide, and what Ignite changed for anyone building agents
Microsoft now ships its own frontier-adjacent coding model inside the subscription it already sells, and starts metering agent run-time in Foundry on …

macOS 27.3 tightens Full Disk Access because of AI agents: silent denials, no one-click toggle, and a 'Private Folders' permission in the code. What agent developers should do before December
Apple is removing the shortcut that let an assistant app earn blanket access to Mail, Messages and backups with a single switch, and the scoped altern…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.