DEV Community

Cover image for llama.cpp 1.0 Ships: Stable API, FP8 Cache, Server Lockdown
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

llama.cpp 1.0 Ships: Stable API, FP8 Cache, Server Lockdown

Verdict: llama.cpp 1.0.0 is a maintenance release for you and a contract for everyone building on it. On November 13, 2026, the ggml-org team at Hugging Face shipped llama.cpp and ggml v1.0.0 — the first stable, semver-governed versions after 3.5 years of development. The version number changes how breakage is signalled to the hundreds of apps that embed these libraries; the code changes that matter are a memory saver for large MoE models, an FP8 KV cache, a new local text-to-speech endpoint — and two changes that can stop a working llama-server cold on upgrade day. Nothing here requires switching runtimes. It requires the short checklist below before you upgrade.

TL;DR

  • A 3.5-year-old project declares stability: llama.cpp and ggml reach 1.0.0, and breaking C API or ABI changes now force a major-version bump (Hugging Face).
  • Upgrade-day breakers: 35 deprecated API functions removed plus llama_set_warmup and the --defrag-thold flag (PR #30565); llama-server now refuses to expose tools or MCP on a non-loopback address without an API key (PR #32113).
  • Gains: an FP8 KV cache, a host-memory MoE expert cache, an OpenAI-compatible /v1/audio/speech endpoint, and a redesigned web UI.
  • Runtime take: Ollama and LM Studio wrap llama.cpp, so your runtime choice does not change at 1.0 — the checklist is what changes.

Why a version number is the headline

llama.cpp runs multi-modal LLMs on everything from a laptop to an enterprise server, and every new integration locks more downstream applications to its interfaces. The team's own reasoning: with local AI and privacy growing, a "flexible but stable" API is "paramount," so v1.0.0 fixes the semver rules all ggml-org projects follow from now on (Hugging Face):

Bump What it signals
MAJOR Breaking change to the public C API in ./include
MINOR Backward-compatible feature, backend update, new model architectures, API addition
PATCH Bug fix with no API change

The version trail staged this deliberately: v0.1.0, then a v0.1.2 pre-release, then v0.9.0 on November 4, 2026, then v1.0.0 (Neurohive).

One boundary matters when you read the promise: the 1.0 contract covers the public C API in ./include only. llama-server command-line flags, the web UI, and the desktop app sit outside it (eesel.ai).

What breaks on upgrade day — and the fix

Dead flags and dead functions. PR #30565 removes llama_set_warmup, 35 deprecated functions, and defrag_thold along with its --defrag-thold command-line flag; an unknown flag stops the tool with an invalid-argument error, so delete it from launch scripts, launchd plists, and aliases (modelfit.io). The NextN/speculative API moves from llama-ext.h into llama.h, and llama_set_embd_state_src is renamed to llama_set_ctx_state_src.

Exposed tools without a key. Starting llama-server with tools or MCP enabled on a non-loopback --host and no API key is now refused at startup. Add --api-key, or opt in explicitly with --allow-unauthenticated-tools for networks you actually trust (PR #32113). This is not paperwork: the built-in web-UI tools include read_file, exec_shell_command, and write_file, which is real exposure on a shared network (eesel.ai). The maintainer merged both security PRs on November 12 and described the trade-off as "hurt the UX a bit" — no CVEs were assigned (The Injection).

SSRF and media paths. The CORS proxy and chat-media URLs could previously resolve to loopback or LAN hosts — server-side request forgery. Version 1.0.0 blocks private targets unless you pass --media-allow-private-urls; a reviewer confirmed a redirect to 127.0.0.1 and a DNS-rebinding host are both refused, and CORS origin matching now checks the port as well as the host (The Injection).

Image and audio restructuring. Image and audio handling is consolidated into a single mtmd component, and MCP tool calls are cancelled on client disconnect or timeout, so integrations that touched the old paths need adapting (Neurohive).

What actually got faster and smarter

The headline memory feature targets mixture-of-experts models: GGUF main weights can stay in system RAM while frequently used experts are held in GPU memory, and when the GPU cache is full, part of the computation falls back to the CPU instead of transferring data — results depend on your hardware and settings (Neurohive). The KV cache now supports OCP FP8, and the release adds support for the Kimi-K3.1-Pro, Ripple-65B-Preview, and xing4_0 text architectures, the drama-3 TTS model, and MageVL-Cowork and Kimi-K3.1 vision models — llama.cpp tracks more than 150 architecture variants in total (GitHub).

ggml 1.0.0 carries the performance work: few-row quantized matmuls and MoE fusion on CUDA and Metal, a Vulkan FlashAttention prefill path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, and expanded Hexagon op coverage. ggml's backend design is what lets the same GGUF file run on virtually any hardware (GitHub). On the server side there is a new OpenAI-compatible /v1/audio/speech TTS endpoint, and the web UI gains chat groups, a live markdown preview, and PDF attachments rendered as text with page thumbnails (Neurohive) — if you run local voice pipelines, this pairs with what we found in our open-source TTS comparison.

Upgrade checklist

  1. Update the binary from the v1.0.0 tag: Homebrew on macOS, Winget on Windows, Docker pull, conda-forge, or build from source. Packages ship across Linux, macOS, Windows, Android, and iOS channels.
  2. Strip --defrag-thold / -dt and any llama_set_warmup call from every launch script before the first run (modelfit.io).
  3. If llama-server exposes tools or MCP beyond loopback, add --api-key before upgrading; use --allow-unauthenticated-tools only on networks you control (PR #32113).
  4. If chat flows load media from a NAS or other LAN address, add --media-allow-private-urls, or move those files locally.
  5. C/C++ embedders: rebuild against the new headers, rename llama_set_ctx_state_src calls, and re-check any use of the 35 removed functions (GitHub).
  6. Point one MoE workload at the host-memory expert cache and test /v1/audio/speech if you serve local speech.

What to run instead: nothing yet

Ollama and LM Studio — the two most popular desktop local runners — are built around llama.cpp and its GGUF format, which is exactly why the semver contract matters to them and not just to C developers. The release changes your runtime decision in one way only: after upgrading, check any start script against the breaking changes above. For most users the honest comparison stays where we left it in our Ollama vs LM Studio verdict: the wrapper you like is still the wrapper you should use, and llama.cpp 1.0 underneath it is what makes that choice safe to keep. If you are sizing hardware for big MoE models, our Edge0-145B open-weights verdict shows what the new expert cache is competing against.

FAQ

Q: Will llama.cpp 1.0.0 break my setup?
A: Only if you use the removed surface: --defrag-thold, llama_set_warmup, the 35 deprecated functions, or unauthenticated tools on a network bind (GitHub).

Q: Why jump from 0.9.0 straight to 1.0.0?
A: The version encodes a promise, not a feature count — from now on, breaking API or ABI changes force a 2.0.0 bump, announced on the semver rules published November 13, 2026.

Q: Does this change Ollama or LM Studio?
A: Not immediately; both wrap llama.cpp, so they inherit the engine changes on their own release schedules while the stable ABI makes those upgrades safer (Hugging Face).

Q: What does the 1.0 label not cover?
A: The public C API in ./include is the contract; llama-server CLI flags, the web UI, and desktop apps can still change in minor releases (eesel.ai).


AI disclosure: this article was drafted with AI assistance and verified against the primary release sources linked throughout. Last verified November 15, 2026. Corrections: how we work.

Top comments (0)