llama.cpp, which runs AI models on local devices and servers, has released a new major version. Developers integrating it can use version numbers to spot integration changes and adopt new model support and server security improvements.

1. Major versions mark compatibility boundaries

On November 13, 2026, the team released v1.0.0 of llama.cpp and its underlying compute library, ggml. Breaking compatibility with the public C API or compiled programs will now raise the first version number to flag the boundary.

The version rules have three tiers. Major releases break the public C API and mark compatibility changes. Minor releases add features or model support without breaking existing usage. Patches fix bugs without API changes or new features. These numbers signal whether integration code may need changes before users read each release's changelog.

2. Public API cleanup and new inference features

The release removes 35 deprecated functions and warmup functions, which prepare the first run, from public headers. It also removes the threshold setting for memory fragmentation, where memory splits into small pieces. The API for speculative decoding, which proposes and checks upcoming outputs, moves from an auxiliary header to the main header. Products calling existing APIs directly can start by checking for removed functions.

Mixture-of-experts models, which select among specialized processing components, gain a cache that holds expert components in main memory. Users can adjust cache size and the CPU's share of processing when a needed expert is missing from the cache. This expands the options for allocating model execution resources.

3. Broader model and hardware support

The release adds support for three text model architectures and speech synthesis models. New text models include Kimi-K3.1-Pro and Ripple-65B-Preview. The team says it now supports over 150 model architecture variants.

Audio and image processing support also expands. The release supports streaming audio input for Gemma 4 and adds image-capable models such as MageVL-Cowork. These changes can benefit products combining text chat with audio and images.

Backends that handle hardware-specific execution can be built as dynamic libraries, shared code files loaded at runtime, and loaded as needed. This release also improves compute paths across several backends, including quantized matrix operations in CUDA and Metal, which compute matrices at lower precision. Neither source provides performance measurements to compare the gains.

4. Server security and workflow changes

llama-server fixes issues including abuse that makes the server send requests to unintended addresses and direct browser access to child instances. Settings now require explicit permission for unauthenticated tool use and media access at private addresses, which are not directly reachable from the public internet.

MCP calls, which use a protocol for connecting external tools, are canceled when the client disconnects or a timeout occurs. The tool server is also notified.

The server's sleep behavior also changes. It will not enter sleep mode while a request is using server state.

The server adds an OpenAI-compatible speech synthesis API.

The web interface lets users organize conversations into named groups and edit long messages with a preview of Markdown, a syntax that uses symbols to format documents.

5. Integration stability and migration costs

The version rules and broader support help maintain integrations and adopt new models. In return, this migration requires checking removed or renamed APIs, plus model files and loading paths as loading requirements tighten for some image models.

Users tuning execution settings can also store the key-value cache in an 8-bit floating-point format. The release notes list the relevant options as follows.

-ctk/-ctv f8_e4m3

The blog outlines prerelease builds for major desktop and mobile operating systems, plus container and package distribution channels. Users can check the channel matching their existing installation method. Those integrating APIs directly can include call-site checks in their upgrade procedure.

The core of v1.0.0 is expanded functionality paired with version numbers that expose compatibility changes, giving integration developers a basis for judging upgrade impact.