llama.cpp 1.0: what changes if you run llama-server
llama.cpp 1.0 is out: a stricter llama-server on your network, an FP8 KV cache option, and removed flags to check before you upgrade.
· 3 min read
llama.cpp 1.0 is out. On 13 November 2026 the ggml team released llama.cpp v1.0.0, with libllama and libggml both at 1.0. For anyone who runs llama-server on their own machines, three things change: a stricter server that refuses some network setups it used to accept, a new FP8 option for the KV cache, and a cleanup that breaks old flags and integrations.
This brief covers what changed, what to check before you upgrade, and what it means if you serve models to other computers on your network.
What happened
The announcement on the Hugging Face blog (13 November 2026) presents 1.0 as a stable base for the software built on top of llama.cpp, and sets out a semantic versioning policy: a major version for a breaking change to the public C API, a minor version for features, backends and model support, and a patch version for bug fixes. The same post says llama.cpp now supports more than 150 architecture variants.
The release also brings new models, including Kimi-K3.1-Pro (#31779) and Ricursive's Ripple-65B-Preview coding model (#31692); what it needs and how fast it runs are in running Ripple-65B locally.
Why it matters: the server is stricter on your network
llama-server still listens on 127.0.0.1 by default, according to the server README. The change is for people who bind it to 0.0.0.0 or a LAN address to reach it from another computer:
- Tools and MCP need a key. With server tools or MCP enabled on a non-loopback address and no
--api-key, the server now refuses to start. You either set a key or pass--allow-unauthenticated-tools(#32113, merged 12 November 2026). - No private media addresses by default. Image and media URLs that point at local or private addresses (a NAS, another machine on the LAN) are refused unless you pass
--media-allow-private-urls. The CORS proxy now reaches public addresses only, so reaching a local MCP server through it no longer works (#32115, merged 12 November 2026).
What it means if you run models yourself
FP8 KV cache. You can now set -ctk f8_e4m3 and -ctv f8_e4m3 (listed as allowed cache types in the server README, where the default is f16). FP8 stores each cached value in 8 bits instead of 16, which is how long contexts fit in less memory. Treat it as experimental for now. The pull request (#28392) publishes no memory, speed or quality figures and no list of backends; CUDA's OCP FP8 support is a separate change (#28413). Compare it with q8_0 on your own prompts before relying on it. For how cache size adds to a model's footprint, see best local LLM by memory.
Ripple and Kimi in plain llama.cpp. Both architectures are now in a stable release, not just on master. Ollama added Ripple in its own 0.40.16 release, as we covered in the Qwen3.8 Flash Next brief.
Apps that bundle llama.cpp will move to 1.0 on their own schedules. Upgrading them means the same flag and API changes, handled by their maintainers.
Sharing llama-server without opening it to the network
If you bound llama-server to 0.0.0.0 only so that a laptop could reach the desktop, the new defaults are a good moment to stop. Corewire's own llama.cpp runners listen on loopback only and are reached from your other devices through your private, encrypted Corewire network, so nothing on the machine is exposed to the LAN. Coding agents reach those models by one name, as the Claude Code guide shows; Corewire is on the download page.