ggml-org, the developer of the open-source LLM inference engine llama.cpp, published v1.0.0, the project's first major version, on its GitHub release page on November 13, 2026. The underlying tensor library, ggml, moved to v1.0.0 at the same time, and the project now commits to signaling breaking changes through major-version bumps. The release also brings new inference features, including FP8 KV caches and a cache for Mixture of Experts (MoE) models.

The Public API Is Cleaned Up, With 35 Deprecated Functions Removed

The centerpiece of v1.0.0 is a cleanup of the public C API in llama.h. According to the release notes, 35 deprecated functions and llama_set_warmup have been removed, along with defrag_thold (--defrag-thold on the command line).

At the same time, the NextN API for speculative decoding, previously in the separate header llama-ext.h, has been promoted into llama.h. llama_set_embd_state_src has been renamed llama_set_ctx_state_src. Apps that embed llama.cpp as a library will need to account for these changes when they update.

Only Breaking Changes Will Bump the Major Version From Now On

In its official blog post on Hugging Face, ggml-org laid out a numbering scheme based on semantic versioning:

  • MAJOR: Breaking change to the public C API in ./include
  • MINOR: Backward-compatible feature, backend updates, model support, or API addition
  • PATCH: Bug fix with no API change

The post states: "Moving forward, both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps." Until now, llama.cpp has frequently published builds numbered with a "b" prefix. Nightly builds still appear on the releases page, so day-to-day updates and milestone releases will be published in parallel.

FP8 KV Caches and an MoE Expert Cache Are Added

On the inference side, the KV cache can now be stored in the OCP FP8 format by passing f8_e4m3 to the -ctk and -ctv options. The KV cache is a memory region that grows with longer context, so storing it in a format with fewer bits reduces the memory required.

For MoE models, a new cache keeps expert weights in host memory. Users can set the share of cache misses handled on the CPU, and LRFU and ClassLRU are available as replacement policies. Two fields, moe_cache_size and moe_cache_cpu_share, have been added to llama_context_params.

On the ggml side, v1.0.0 removes deprecated APIs and adds YaRN truncation and FP8 support.

New Models, a Redesigned Web UI, and Server Security Fixes

Newly supported models are Kimi-K3.1-Pro, Ripple-65B-Preview (ricursive), xing4_0, and the speech synthesis model drama-3 TTS (fishaudio). Through mtmd, the multimodal component, image input is now supported for MageVL-Cowork and Kimi-K3.1. The server also gains a /v1/audio/speech endpoint for speech synthesis.

The bundled web UI has been redesigned. It adds sidebar folding and groups, a markdown editor with live preview, a mobile layout, right-click menus, and thumbnails for PDF attachments.

On server security, the release introduces two options, --allow-unauthenticated-tools and --media-allow-private-urls, and includes fixes related to CORS, DNS rebinding, and SSRF.

Our Take

llama.cpp is widely used to run LLMs on local PCs, and it is often embedded as the inference engine inside other apps. In our view, the real significance of this release is that the written versioning rules make it easier for those integrators to gauge the impact of an update.

That said, v1.0.0 itself removes deprecated APIs, so apps that still rely on old functions will face migration work. How often the major version will change remains to be seen as further releases arrive.

Summary

llama.cpp and ggml both reached v1.0.0 on November 13, cleaning up the public API and establishing rules that signal compatibility through version numbers. The feature changes are broad as well, covering FP8 KV caches, a host-memory MoE expert cache, new model support, a redesigned web UI, and server security fixes. Full details are available in the GitHub release notes and on ggml-org's blog.