On November 13 the ggml-org team released llama.cpp and ggml 1.0.0. Many of the tools for running language models locally are built on them. The headline is not speed but a promise: from now on, any change that breaks the C API means a new major version.
llama.cpp and ggml 1.0: semantic versioning, an API cleanup and a safer server
ggml is a tensor library that does the math on CPUs and GPUs. llama.cpp builds on it to run language models and adds the libllama library, the llama cli program and an HTTP server. ggml v1.0.0 went up on GitHub first, and llama.cpp v1.0.0 followed less than two hours later. The llama.cpp repository dates back to March 2023 and has more than 135,000 stars.
What the 1.0 means
In a post on the Hugging Face blog, the team explains that both libraries now follow semantic versioning. The major version goes up when a change breaks the public C API (the headers in the include directory), the minor version for new features, backends and model support, and the patch version for bug fixes only. A major bump also signals an ABI break, meaning programs linked against the library must be recompiled.
ggml shipped 0.x releases before, and llama.cpp has used vX.Y.Z release tags since August. According to the release and versioning notes for ggml-org projects, llama.h is the public API, while the helper header llama-ext.h can change without a version bump. Nightly builds with tags like b12359 will keep coming.
Cleaning up before making the promise
Promising stability meant clearing out old code first. llama.h loses 35 deprecated functions, along with llama_set_warmup and the --defrag-thold option. The NextN API for speculative decoding moves from llama-ext.h into the public header. On the ggml side, the legacy DIAG_MASK ops are gone and the RPC protocol moves to version 9, so nodes running an older protocol are rejected when they connect. If you write C or C++ against these libraries, check for removed calls before upgrading.
New features and a safer server
llama.cpp gains FP8 KV caches, a host-memory cache for MoE experts with a configurable CPU share, and support for new architectures, including Kimi-K3.1-Pro and Ripple-65B-Preview. The blog post counts more than 150 supported model architecture variants. The server has a new OpenAI-compatible /v1/audio/speech endpoint, and the ggml backends, from CUDA to WebGPU, got performance work.
If you expose the server on a network, the security changes matter more. The release closes an SSRF hole in the CORS proxy, blocks DNS rebinding and fixes the handling of --cors-origins. Serving tools without authentication and fetching media from private addresses are now opt-in (--allow-unauthenticated-tools, --media-allow-private-urls), and the HF_TOKEN is only sent to huggingface.co hosts. In our view, developers who embed libllama in their apps gain the most: they now know when an update might break something.
What this means if you are learning to code
A version number is a contract between a library's authors and its users. When your package.json says ^1.4.2, npm may install 1.5.0 but never 2.0.0, because it assumes a change in the first number can break your code. 0.x versions make no such promise, and by moving llama.cpp to 1.0 its authors are making one. You can practice dependencies and package.json in our JavaScript and TypeScript course.
Two commands from the README are enough to see a local model at work: llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF downloads a small model from Hugging Face and opens a chat in your terminal, and llama serve starts a server with an OpenAI-compatible API. You can then test code written for a cloud API with no key and no token bill. We explain where models come from in our article on Hugging Face, and if you want an easier start, read our guide to Ollama.
A small model will not replace a cloud assistant, but it is enough to understand how an app talks to a model. Working with AI assistants while coding is what our AI-assisted programming course is about. FP8, MoE caches and GPU backends can wait.
Also today
- Rust 1.100. The November 12 release stabilizes the never type
!and the allocator API, for exampleVec::new_in. Cargo now warns about unused dependencies and can skip freshly published crate versions, for example anything from the last 3 days (global-min-publish-age). Source - Node.js security releases. Versions 22.23.5, 24.22.3 and 26.13.1 came out on November 11 and fix 16 vulnerabilities, 7 of them rated high, mostly in HTTP/2 and TLS. If you run a Node.js server, update it within your LTS line. Source
- NestJS 12.5.0. The November 12 release tightens default limits:
WsAdapteraccepts messages up to 1 MB (previously 100 MB), and multipart requests can have at most 1,000 parts. A guard, interceptor or pipe that cannot run now stops startup, so after updating a project from our NestJS course, check that it still boots. Source - AWS Emulator. On November 11 AWS released a free local emulator of its APIs, written in Rust, for the AWS CLI, SDKs and coding agents. Agents write and maintain the code, and IAM policies are not evaluated the way AWS does it, so it is not for security testing. Source
- Apache HTTP Server 2.4.71. The November 11 release fixes CVE-2026-119618, a critical flaw in mod_http2 that can lead to remote code execution. Versions 2.4.55 through 2.4.70 are affected. Source
- Project Zero shortens its deadlines. Google's team said on November 13 that it now discloses bugs 90 days after reporting them, fixed or not. The window drops to 60 days on February 2, 2027 and to 30 on January 3, 2028. Source
- Cloudflare open-sources Kitesurf. The experimental browser engine for AI agents runs on Cloudflare Workers, speaks the Chrome DevTools Protocol (so Puppeteer can connect) and is licensed under Apache 2.0. Source
Frequently asked questions
What is the difference between llama.cpp and ggml?
ggml does the tensor math on a range of hardware, from CPUs to CUDA, Metal and Vulkan. llama.cpp uses ggml and adds what language models need: loading GGUF files, the libllama library, a terminal chat and a server. Both come from ggml-org and share the same versioning rules from 1.0 on.
Will my code break after upgrading to llama.cpp 1.0?
C or C++ code that calls functions from llama.h may need changes, because the release removes 35 deprecated functions. If you use the server over HTTP, check the new security options first, since some behavior is now opt-in. The full list is in the GitHub release notes.
How do I run a model locally with llama.cpp?
Install llama.cpp from a package (Homebrew, winget, Docker) or with the script from the project README. Then llama cli -hf with a Hugging Face repository name downloads the model and opens a chat, and llama serve starts a server compatible with the OpenAI API.
Sources
- llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem, Hugging Face (ggml-org), November 13, 2026
- llama.cpp v1.0.0, ggml-org (GitHub), November 13, 2026
- ggml v1.0.0, ggml-org (GitHub), November 13, 2026
- Releases and versioning of ggml-org projects, ggml-org (GitHub), August 7, 2026
- llama.cpp README, ggml-org (GitHub)
- Announcing Rust 1.100.0, The Rust Release Team, November 12, 2026
- Wednesday, November 11, 2026 Security Releases, Node.js, November 11, 2026
- NestJS v12.5.0, NestJS (GitHub), November 12, 2026
- Test your applications against AWS locally with AWS Emulator, AWS News Blog, November 11, 2026
- Apache HTTP Server 2.4 vulnerabilities, Apache Software Foundation, November 11, 2026
- Policy and Disclosure: 2026 Edition, Google Project Zero, November 13, 2026
- Kitesurf is now open source, Cloudflare, November 13, 2026
This article was prepared with the help of AI from primary sources and reviewed by the CodeWorlds editorial team.