llama.cpp Hits 1.0: Stable Local AI, Free to Run
In short
- llama.cpp and its engine ggml both hit version 1.0.0 on November 13, with real version numbers from now on.
- The release removes old API, adds FP8 support, an audio speech endpoint and a hardened server.
- Money angle: running models on your own machine is the cheapest way to stop paying per seat.
The most popular tool for running AI on your own computer just got its first stable version. llama.cpp is now 1.0.0, and so is ggml, the engine under it. For a builder, that means you can finally depend on it without fearing that the next update breaks everything.
What happened
On November 13 the ggml-org team tagged ggml v1.0.0 and merged the llama.cpp version bump to 1.0.0 about half an hour later. Until now the project used build numbers like b12377. Now it uses semantic versioning: major, minor, patch. A major number change means something broke. A minor one means something new. A patch means a fix.
What is inside llama.cpp 1.0.0, per the release notes:
- Deprecated public API removed, and the speculative decoding API moved into the main header.
- FP8 KV caches, which store the model’s working memory in less space.
- A host-memory cache for mixture-of-experts models, with a CPU share you can set.
- New model architectures, including Kimi-K3.1-Pro, plus a text-to-speech model.
- A new /v1/audio/speech endpoint in llama-server.
- Server hardening against SSRF, CORS abuse and DNS rebinding, and a redesigned web UI with chat groups.
ggml itself adds speedups for CUDA, Metal, Vulkan, SYCL and OpenCL, so the same model should run faster on Nvidia, Apple, AMD, Intel and phone chips.
How to use it
Do the math: a small team of 5 paying €20 per person a month for AI tools spends €100 a month, or €1,200 a year. A local model on a machine you already own costs electricity. That is my own example, not a price from the release. Quality of local models still trails the biggest paid ones, so test before you cancel anything.
Three practical steps:
- Pin your version. Pick 1.x in your deployment and you will not be surprised by a silent breaking change.
- Put llama-server behind your internal tools. It speaks an OpenAI-style API, so an automation in Make or n8n can call it like any other model.
- Use the new speech endpoint for voice notes and read-aloud features without paying per character.
One warning. If you built something on the old API, 1.0.0 removes it. Read the API changes before you update.
What it means
A 1.0 label tells companies this is safe to build products on. Expect more apps to ship with llama.cpp inside. The losers are thin wrappers that charge a monthly fee for something a free engine now does reliably.
Read also: Run a Team of AI Coworkers on Your Own Server With Rerun · DeepSeek Open-Sources an Agent That Answers Your GitHub Issues · Nanbeige4.3-33B: A Free Agent Model You Can Run Yourself
Ivan’s Take
Worth it if you have any repetitive text work: product descriptions, email sorting, internal search. I would test it on a spare machine this week and move the boring tasks off paid APIs first. For hard reasoning I would still pay for a top model.
HawkValue
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
