MONDAY, NOVEMBER 16, 2026 Tech news, minus the noise.
N News Cast
LATEST
Home  /  Dev
Dev

llama.cpp 1.0 Is Here, and Local AI Apps Finally Have a Stable Foundation

llama.cpp 1.0.0 arrived on November 13 with a stable public API, semantic versioning, and FP8 KV caches that halve conversation memory. It is also a breaking release, so check for removed functions first.

llama.cpp 1.0 Is Here, and Local AI Apps Finally Have a Stable Foundation
Source: Wikimedia Commons (CC0)

llama.cpp reached version 1.0.0 on November 13, 2026, and the headline is not a new model or a speed boost. It is a promise. After three and a half years of build-number versioning, the open-source engine behind much of today's local AI software now has a stable public API, and breaking changes will arrive only with a major-version bump.

What llama.cpp is, and why 1.0 matters

llama.cpp is not a language model. It is an inference engine written in C and C++, built on top of the ggml tensor library, that lets ordinary computers run large models locally. If you have used a local AI app, a desktop chatbot, or a self-hosted model server, there is a good chance llama.cpp was doing the work underneath. The project supports more than 150 model-architecture variants, and its libraries ship on Linux, macOS, Windows, Android, and iOS. It is the kind of infrastructure that makes local AI practical in the first place, the same way Rust 1.100 gave the language a long-awaited stable feature after ten years.

Until now, llama.cpp was versioned by build numbers, and the last numbered releases were v0.9.0 for llama.cpp and v0.29.0 for ggml. Downstream apps had no reliable signal for whether an update would break their code. With v1.0.0, both libraries adopt semantic versioning with a refreshingly specific rule: a MAJOR bump means a breaking change to the public C API in the include directory, a MINOR bump is a backward-compatible feature or addition, and a PATCH is a bug fix with no API change.

The cleanup that came with the milestone

A stable contract does not mean 1.0 is a drop-in upgrade. The release uses the major-version milestone to remove the deprecated public API: 35 deprecated functions are gone, along with the llama_set_warmup setter and the defrag-thold option. The NextN speculative-decoding API moves from the extension header into the public llama.h header. If your code calls any of the removed functions, it will break at compile time, which is the point of doing the cleanup now rather than later.

The companion ggml v1.0.0 release does the same kind of pruning on the tensor library side, removing deprecated APIs and legacy operations, adding configurable YaRN truncation and FP8 support, and avoiding uploads of scheduler inputs that are not read. The maintainers added a related safeguard in August: an official llama.cpp release tag is created only when its internal ggml copy is identical to the released ggml copy, so the two libraries cannot drift apart silently.

What is new under the hood

The most practical change for people running models is FP8 KV-cache support through the f8_e4m3 type, with optional static per-tensor scales. The key-value cache is the conversation memory a model keeps while generating, and halving its precision halves its size. For a Qwen2.5-7B model with a 32,768-token context, the full cache drops from 1,792 MiB to 896 MiB, which is the difference between a long conversation fitting in memory and swapping to a crawl.

Version 1.0 also adds a host-memory cache for mixture-of-experts models with a configurable CPU share, support for three new text architectures (Kimi-K3.1-Pro, Ripple, and xing4_0), and server improvements including chat groups, a redesigned browser interface, a live-preview Markdown editor, mobile layout improvements, and an OpenAI-compatible speech-generation endpoint for supported models. Server security fixes are included as well.

What to do before you upgrade

If you run llama.cpp directly, check your launch commands for the removed defrag-thold option and delete it. If you build against the libraries, grep for the 35 removed functions before upgrading, because they fail at compile time now. If you only use llama.cpp through another application, you do not need to do anything: your app's maintainer handles the upgrade, and the version number will tell them whether it is safe.

The real test of 1.0 is the next breaking change. Tooling is having a good month overall: .NET 11 shipped with C# 15, and GitHub rebuilt its Git storage to handle thousands of AI agents pushing at once. If the team holds to its rules, the next removal from the public C API will arrive as v2.0.0 rather than inside a routine update. That is a small thing to promise and a hard thing to keep, and it is exactly what an ecosystem of hundreds of downstream apps needs.

AS
News Cast Editorial Team

We cover AI, gadgets, startups and security — original reporting and explainers, written by humans, for humans. Got a tip? Reach us via the contact page.