Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableNeurohive→ original

llama.cpp 1.0.0: Stable API for Developers and 35 Old Functions Removed

llama.cpp, a popular tool for running neural networks on your own computer, has released version 1.0.0 with its first stable API, the set of functions that applications use to talk to the library. From now on, the first digit of the version number tells you whether your app will break after an update: it changes only on incompatible changes. The upgrade is not free, though. The team removed 35 deprecated functions, so developers should check their builds in advance.

AI-processed from Neurohive; edited by Hamidun News
llama.cpp 1.0.0: Stable API for Developers and 35 Old Functions Removed
Source: Neurohive. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceNeurohive ↗
Editor-in-Chief
Large MoE modelspart of the compute can be offloaded to the CPU, not just the GPU
Key metric #2The server gained an OpenAI-compatible request for text-to-speech
Key metric #3The Hugging Face token now goes only to huggingface.co by default

Imagine you built an AI chat that runs a local model through llama.cpp. You updated the library, and the chat stopped launching: the developers renamed or removed a function you were calling. For a project that underpins a large share of local AI services, surprises like this were routine. Now llama.cpp has reached version 1.0.0 and for the first time formalized a compatibility promise. This is not a speed leap but a sign that the infrastructure has matured. Alongside it, ggml, the component that performs the math, also moved to version 1.0.0.

What a "stable API" means and why the version number matters

An API is the set of functions and rules through which your application talks to a library. An ABI is the same thing at the level of the compiled binary: if it changes, even unchanged code has to be recompiled. llama.cpp had version numbers before, but they came with no hard guarantees. Starting with 1.0.0, it follows what is known as semantic versioning.

  • The first digit goes up when the API or ABI changes incompatibly. Moving from 1.0.0 to 2.0.0 means your code will probably need fixing.
  • The second digit marks compatible additions, that is, new features that break nothing.
  • The third digit signals bug fixes that leave the interface unchanged.

It works like a power socket standard: as long as the standard doesn't change, any device fits. App authors can update freely within the 1.x line and prepare for a major version change and test it in advance. Projects ship two libraries: libllama and libggml.

This is also a plus for Linux distributions. According to a member of the Debian community, predictable compatibility makes llama.cpp easier to package. But the project has eight distribution channels, from ready-made files on GitHub to software installation catalogs and Apple app bundles. Whether 1.0.0 has actually arrived in each of them has to be checked separately.

The price of the upgrade: 35 removed functions

Moving to 1.0.0 may itself require changes. The team removed 35 deprecated functions, including the function for configuring model "warm-up", that is, the preparatory computation before the first response. In addition, part of the image and audio code was restructured and merged into a single component. Applications that used these capabilities will need to be adapted.

Among the memory-related changes: more compact storage for intermediate computations was added. The launch of MoE models has also changed, where only a subset of "experts" is activated for each token. The main data for such models can reside in system RAM, while frequently used experts are additionally cached in GPU memory. If the needed expert is not in that cache, once the cache is full, part of the computation can be handed to the CPU rather than transferring data to the GPU. This may speed things up, but the outcome depends on the hardware and settings.

The list of supported models now includes Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0 for text, drama-3 for speech generation, and MageVL-Cowork and Kimi-K3.1 for image processing. In total, according to the team, the tool supports more than 150 architectures.

Voice, browser and server protection

The llama.cpp server, which accepts requests from a browser or app and passes them to the model, gained an OpenAI-compatible way to request text-to-speech. If your application already sends requests in the OpenAI format, wiring up local speech generation becomes easier. The browser interface received chat groups, a preview of message formatting, and an improved layout on phones.

Access protection for home or work networks has also been strengthened:

  • running tools on a server reachable from other devices without an API key now requires explicit permission;
  • loading images and other media from the internal network also has additional restrictions;
  • the Hugging Face access key is no longer sent automatically to arbitrary model download addresses. By default it goes only to huggingface.co, and any other trusted address has to be allowed explicitly.

What to do in practice

App authors should build their projects against 1.0.0 and check whether removed functions are being called, especially model "warm-up" and the image and audio code. Those running a server for a team should review their access settings: without an API key, tools will now not run unless explicitly permitted.

There are limits too. The release does not guarantee a speedup for any given model, and the team has not published a general performance comparison with the previous version. New capabilities will only reach users as the developers of finished applications update them. The main takeaway of 1.0.0 lies elsewhere: local AI now has a predictable foundation, and products built on it can be planned months ahead.

Story

Key Questions Answered

What a "stable API" means and why the version number matters?

An API is the set of functions and rules through which your application talks to a library. An ABI is the same thing at the level of the compiled binary: if it changes, even unchanged code has to be recompiled. llama.cpp had version numbers before, but they came with no hard guarantees. Starting with 1.0.0, it follows what is known as semantic versioning. The first digit goes up when the API or ABI changes incompatibly. Moving from 1.0.0 to 2.0.0 means your code will probably need fixing.

What are the key details on The price of the upgrade: 35 removed functions?

Moving to 1.0.0 may itself require changes. The team removed 35 deprecated functions, including the function for configuring model "warm-up", that is, the preparatory computation before the first response. In addition, part of the image and audio code was restructured and merged into a single component. Applications that used these capabilities will need to be adapted. Among the memory-related changes: more compact storage for intermediate computations was added.

What are the key details on Voice, browser and server protection?

The llama.cpp server, which accepts requests from a browser or app and passes them to the model, gained an OpenAI-compatible way to request text-to-speech. If your application already sends requests in the OpenAI format, wiring up local speech generation becomes easier. The browser interface received chat groups, a preview of message formatting, and an improved layout on phones.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…