llama.cpp Hits 1.0.0: Breaking the C API Now Requires a Major Version Bump
llama.cpp, the popular engine for running LLMs on your own hardware, has shipped version 1.0.0 with a simple rule: if you break the public C API in ./include, bump the first number (major) to 2.0. Until now, projects that embed the library have been stuck on nightly builds, worried that their code would stop compiling after an update. Now they can pin to 1.0.0 and move forward deliberately.
AI-processed from X: @Luejinlynn; edited by Hamidun News
Imagine you've embedded llama.cpp in your application, then a week later you pull the latest build and the project no longer compiles, because the library's author renamed or removed a function. That's how many people who use this engine to run LLMs on their own hardware have lived: there were no promises of interface stability, so they had to track nightly builds straight from development. On November 13, 2026, that changed. llama.cpp shipped version 1.0.0, and the same day the ggml library it's built on also reached v1.0.0.
What the "1.0.0 contract" means in plain terms
The main value of this release isn't another flag, it's a rule of the game. The official Hugging Face blog states that if the public C API in the ./include directory breaks (that is, the functions that other programs call), the MAJOR version is incremented, the first number. New backward-compatible features and model support bump MINOR, the second number. Bug fixes alone are PATCH, the third number.
In practice, this means that if your dependency is pinned at 1.0.0, upgrading to 1.1 or 1.2 should not break your code. But the appearance of 2.0 is a signal to review the changelog and fix your calls. Linux distributions, package managers, and applications linking against the library can finally pin the version instead of drifting along with nightly.
Caveat: v1.0 is a contract, not a promise that "the interface will never change." If the authors need a breaking change, they will release 2.0. That's why you need to watch the major version specifically.
What else is in the release
- Deprecated public APIs have been removed, and the NextN and speculative decoding (a method for accelerating generation) interfaces have been promoted into the main header, llama.h.
- FP8 KV cache has been added. The KV cache stores intermediate context data, and a more compact format helps fit long conversations into the same memory. It's enabled with the -ctk and -ctv flags set to f8_e4m3.
- For MoE models, there's now an expert cache: experts that don't fit on the GPU can be kept in host RAM, and the share offloaded to the CPU is configurable. It's useful where VRAM isn't enough to hold all experts at once.
- llama-server gained protection against SSRF (when a server is tricked into reaching internal addresses), CORS attacks, and DNS rebinding. In addition, MCP tool calls are now cancelled if the client disconnects or the timeout expires.
- An OpenAI-compatible /v1/audio/speech endpoint has been added for speech synthesis.
The security section matters to anyone exposing the server externally. When a server can run tools via MCP, it effectively acts on behalf of the user, and these are often points of attack.
What developers should do
The author of the original analysis suggests three steps.
- Pin the version in your build or package manager to 1.0.0 and stop silently following nightly.
- If you serve llama-server to other people, carefully review the new flags --allow-unauthenticated-tools and --media-allow-private-urls. By default, it's better not to enable them: the first allows tool calls without authentication, the second permits requests to private addresses when loading media.
- Before upgrading, run your own API calls and verify that you aren't using removed functions. This is the author's personal advice; they did not test it in practice themselves.
Limitations
The blog mentions support for more than 150 architectures, but that is the authors' own wording. Whether a specific model will work on 1.0.0 should be checked in that version's model list. In addition, information about the release so far comes from official release notes and the blog. Real compatibility issues will only surface once the project has lived under the new rules for a while.
Bottom line: for the local-LLM community, v1.0.0 is first and foremost a sign of process maturity. Teams now have a foundation for planning upgrades, and for a product built on llama.cpp, that matters more than any single flag.
- Nov 21, 2026Linger Runs 12 AI Models on Your PC: Up to 5x Faster and $0 in Token Costs
- Nov 21, 2026QA Engineer Cuts Autotest Development From 2 Weeks to 2 Days Using an LLM
- Nov 21, 2026Hermes Lets You Move a Local Agent to the Cloud in One Step from Desktop
- Nov 21, 2026Apple Overhauls Mac Security in macOS 27.3 to Address AI Agents With Full Disk Access
Key Questions Answered
What the "1.0.0 contract" means in plain terms?
The main value of this release isn't another flag, it's a rule of the game. The official Hugging Face blog states that if the public C API in the ./include directory breaks (that is, the functions that other programs call), the MAJOR version is incremented, the first number. New backward-compatible features and model support bump MINOR, the second number. Bug fixes alone are PATCH, the third number.
What else is in the release?
Deprecated public APIs have been removed, and the NextN and speculative decoding (a method for accelerating generation) interfaces have been promoted into the main header, llama.h. FP8 KV cache has been added. The KV cache stores intermediate context data, and a more compact format helps fit long conversations into the same memory. It's enabled with the -ctk and -ctv flags set to f8_e4m3.
What developers should do?
The author of the original analysis suggests three steps. Pin the version in your build or package manager to 1.0.0 and stop silently following nightly. If you serve llama-server to other people, carefully review the new flags --allow-unauthenticated-tools and --media-allow-private-urls. By default, it's better not to enable them: the first allows tool calls without authentication, the second permits requests to private addresses when loading media.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.