llama.cpp v1.0.0 brings stability to the growing world of local AI
The popular open-source tool that lets anyone run AI models on their own computer has reached version 1.0, making it more reliable for developers building private AI apps.

What you could do with it
- A developer building private AI apps could rely on a stable C/C++ API that guarantees major-version bumps before breaking changes, making their integration reliable long-term 1.
- A researcher could run over 150 model architectures—including text, vision, and speech models—on everyday laptops or high-end servers without sending data to cloud providers 1.
- An app maker could add voice generation to their software through a new OpenAI-compatible audio endpoint, requesting TTS output via a familiar interface 2.
A stable foundation for local AI
The team behind llama.cpp said the project started 3.5 years ago and has grown into something many other applications depend on 1. Today they released libraries llama.cpp v1.0.0 and ggml v1.0.0, marking a shift toward long-term stability 2.
Moving forward, any change that breaks compatibility with older versions will require a new major version number. "Having a flexible but stable API is paramount," the team wrote in a blog post on Hugging Face, noting that every new integration means more developers relying on their interfaces 1.
The underlying engine, called ggml, is designed to run on virtually any hardware — from regular laptops to high-end servers — by using dedicated backend modules that can be compiled separately or loaded while the program is running 12.
What llama.cpp brings to local AI
- Cloud AISends your data to big companies
- 150+ model typesText, vision, and speech models
- Any hardwareWorks on laptops and servers
- Your data stays privateNo need to send conversations away
From the story: local vs cloud, model count, hardware range, and privacy benefit
New models and features
This release adds support for several new architectures. Three text models are now supported: Kimi-K3.1-Pro, a dense 65-billion-parameter model called Ripple-65B-Preview, and xing4_0 2. Vision and speech capabilities have also been expanded with MageVL-Cowork, Kimi-K3.1 vision, and the drama-3 text-to-speech system . Overall, llama.cpp now supports more than 150 model architecture variants 1.
Performance improvements include FP8 key-value caches, which reduce memory use during inference, and a new method for handling Mixture of Experts models that allows them to share CPU memory when needed 2. The update also adds an OpenAI-compatible audio endpoint, letting programs request voice generation through a standard interface .
Security was strengthened too. The built-in server component received hardening against several common web attacks, including SSRF (where a malicious site tricks a server into contacting internal resources), DNS rebinding, and cross-origin issues 2.
From the official announcement
Who can use it
llama.cpp is distributed through multiple channels to reach different users. It is available as Docker images for Linux, Debian packages, Homebrew and conda-forge for macOS and Linux, MacPorts for macOS, Winget for Windows, and XCFrameworks for iOS and macOS 1. Pre-release builds are also offered for Android and iOS .
The project says it continues to rely on an open-source community of contributors and has invested in better testing infrastructure, using hardware from companies including NVIDIA, AMD, Qualcomm, Arm, IBM and Intel 1.Llama Desktop apps for MacOS and Windows, shipped by the llama.cpp and Hugging Face teams, are also mentioned as a way for everyday users to try local AI without setting up code themselves .
This story reports on a trend for information only. It is not advice or encouragement to do any of it.
Corrections and updates
- Corrected on 14 November 2026. An earlier version included an image of an open ancient book under warm candlelight; neither source mentions such an image, so the replacement removes unsupported content.
Key facts
- llama.cpp v1.0.0 and ggml v1.0.0 were released together, establishing strict semantic versioning rules
- Any breaking change to the public API will now require a major version bump
- Support added for Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0, drama-3 TTS, and MageVL-Cowork vision
- Over 150 model architecture variants are now supported
- Distributed via Docker, Homebrew, conda-forge, Winget, MacPorts, Debian, and XCFramework
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
New modelsCompressed 27B Model Beats Full Version at Tool Calling in Under 8 GBHerald of AI
Compressed 27B Model Beats Full Version at Tool Calling in Under 8 GB
A heavily compressed version of Qwen3.8-27B runs on everyday hardware and outperforms the original on agent skills.
Apps and productsYogurt lets you record and summarize meetings with AI that runs entirely on your own computerHerald of AI
Yogurt lets you record and summarize meetings with AI that runs entirely on your own computer
A free open-source app turns your calls into transcripts and summaries without sending any data to the cloud.
CoreWeave goes all-in on open-source Kubernetes as AI model launches get harder
It is getting harder to run brand-new AI models on company servers right after they launch — and a growing cloud provider is betting that fixing that will matter to everyone who uses AI tools.