# llama.cpp reaches v1.0.0 with semantic versioning, FP8 KV caches and a MoE expert cache

> OnPrem World · https://onpremworld.com/en/news/2026-11-llama-cpp-v1-0-0 · CC BY 4.0

Published: 2026-11-13 · Runtimes & apps

llama.cpp and its ggml library both reached version 1.0.0 on November 13, and from now on a major version bump will signal any breaking change to their public C API. The release removes 35 deprecated functions, adds FP8 (E4M3) KV caches, and adds a host-memory cache for MoE experts with a configurable share of work left on the CPU. It supports three new text architectures, Kimi-K3.1-Pro, Ripple and xing4_0, plus drama-3 text-to-speech and Kimi-K3.1 and MageVL-Cowork vision. `llama-server` gets security fixes for CORS, DNS rebinding and SSRF, an OpenAI-compatible `/v1/audio/speech` endpoint and a redesigned web UI with chat groups.

## Why it matters locally

The engine underneath many local AI apps now promises stable versioning, and its FP8 KV cache and MoE expert cache aim to fit longer contexts and larger MoE models into the memory you already have.

## For the market

A 1.0 label with a semantic-versioning promise makes llama.cpp easier to justify inside companies that need predictable upgrades, which likely helps the apps built on it more than llama.cpp itself. The v1.0 announcement credits hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel, a sign that chip vendors now treat it as a target worth backing.

## In plain words

llama.cpp is the free engine many "run AI on your own computer" apps use under the hood. Version 1.0 is its official grown-up release: fewer surprise breakages, and it can squeeze bigger models into the same computer.

## For practitioners

Try `-ctk f8_e4m3 -ctv f8_e4m3` to cut KV cache memory for long contexts, and the new `moe_cache_size` / `moe_cache_cpu_share` context parameters when MoE experts don't fit in GPU memory; measure quality and speed on your own prompts first. Code linking `libllama` directly must migrate: 35 deprecated functions, `llama_set_warmup` and `--defrag-thold` are gone, and `clip.h` is folded into `mtmd`. If you expose `llama-server` on a network, note the new `--allow-unauthenticated-tools` and `--media-allow-private-urls` opt-ins.

## Source

[llama.cpp v1.0.0 release notes (GitHub)](https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0)

[All news](https://onpremworld.com/en/news.md)

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.
