llama.cpp 1.0: what breaks and how to upgrade

On Friday 13 November, llama.cpp tagged v1.0.0, together with v1.0.0 of ggml, the tensor library underneath it. From this release the project follows semantic versioning: a major version bump means a breaking change to the public C API. To get to a clean starting point, 1.0 removes 36 functions that were already marked deprecated, drops one command-line option, and makes llama-server stricter about what it will reach and who may call it.

If you only run the binaries, most of this is a change of defaults. If your code or your bindings call llama.h, you may have a build to fix. Here is what changed, what replaces it, and how to check your own code before you upgrade.

What 1.0 means

The launch post, published by the llama.cpp team at Hugging Face, sets out the rules for every ggml-org project that uses semantic versioning:

Bump	When
Major	"Breaking change to the public C API in ./include"
Minor	"Backward-compatible feature, backend updates, model support, or API addition"
Patch	"Bug fix with no API change"

The post adds that "both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps." It does not say how long the 1.x line will be supported, and it lists no migration steps. Those are in the pull requests, covered below.

Two details of the release process matter in practice. Both come from the project's release and versioning document, which the release notes link:

• The numbered nightly builds continue. Tags such as b12377 are development builds from master, cut several times a day. The vX.Y.Z tags are the official releases, and a release tag is only created when llama.cpp's internal copy of ggml is identical to the released ggml. That guarantee is what lets a package manager link llama.cpp against a system-wide ggml.
• Not every channel ships the release. The document says Homebrew builds from the release tag, while winget and the project's own llama.app installer publish the nightly builds. On 14 November, Homebrew's formula listed 1.0.0 as its stable version.

The v1.0.0 release page itself carries no binaries. Its only asset is a text file, and the notes point to nightly build b12359 for downloads.

To build the release yourself, check out the tag and turn off the development flag, which is on by default and appends -dev to the version:

git clone --branch v1.0.0 --depth 1 https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build -DLLAMA_BUILD_IS_DEV=OFF
cmake --build build --config Release

What breaks if you call the C API

One pull request, #30565, did the pruning. It removed llama_set_warmup and all 35 other functions marked DEPRECATED in llama.h. The replacements below are the ones named in the deprecation messages that the pull request deleted.

Removed in 1.0	Use instead
llama_load_model_from_file	llama_model_load_from_file
llama_free_model	llama_model_free
llama_new_context_with_model	llama_init_from_model
llama_n_ctx_train, llama_n_embd, llama_n_layer, llama_n_head	llama_model_n_ctx_train, llama_model_n_embd, llama_model_n_layer, llama_model_n_head
llama_n_vocab	llama_vocab_n_tokens
llama_get_state_size	llama_state_get_size
llama_copy_state_data	llama_state_get_data
llama_set_state_data	llama_state_set_data
llama_load_session_file, llama_save_session_file	llama_state_load_file, llama_state_save_file
llama_token_get_text, llama_token_get_score, llama_token_get_attr	llama_vocab_get_text, llama_vocab_get_score, llama_vocab_get_attr
llama_token_is_eog, llama_token_is_control	llama_vocab_is_eog, llama_vocab_is_control
llama_token_bos, _eos, _eot, _sep, _nl, _pad	llama_vocab_bos, _eos, _eot, _sep, _nl, _pad
llama_token_fim_pre, _suf, _mid, _pad, _rep, _sep	llama_vocab_fim_pre, _suf, _mid, _pad, _rep, _sep
llama_token_cls, llama_vocab_cls	llama_vocab_bos (the header noted that CLS is equivalent to BOS)
llama_add_bos_token, llama_add_eos_token	llama_vocab_get_add_bos, llama_vocab_get_add_eos
llama_sampler_init_grammar_lazy	llama_sampler_init_grammar_lazy_patterns
llama_set_warmup	No replacement. The deprecation note said user code should do warmup runs manually

Four more changes can break a build or a script:

• defrag_thold is gone from llama_context_params. Removing a field changes the layout of the struct, so anything compiled against an older header must be recompiled, not just relinked.
• The -dt, --defrag-thold option is gone. Before 1.0 it did nothing except print a deprecation warning. Remove it from launch scripts and service files.
• clip.h is removed. The release notes say its model code was consolidated into mtmd, the multimodal library.
• One rename in the speculative decoding API. 18 functions for NextN and draft models moved from the internal llama-ext.h header into the public llama.h, and llama_set_embd_state_src became llama_set_ctx_state_src.

To find the removed names in your own code, search for them before you upgrade:

grep -rnE 'llama_(load_model_from_file|free_model|new_context_with_model|n_(ctx_train|embd|layer|head|vocab)|get_state_size|copy_state_data|set_state_data|(load|save)_session_file|set_warmup|token_(get_(text|score|attr)|is_(eog|control)|bos|eos|eot|cls|sep|nl|pad|fim_(pre|suf|mid|pad|rep|sep))|add_(bos|eos)_token|vocab_cls|sampler_init_grammar_lazy)\(' src/

The pattern matches all 36 removed functions and skips the current names that look similar, such as llama_token_to_piece and llama_sampler_init_grammar_lazy_patterns. If you compile with warnings on, you have probably seen most of these flagged as deprecated already.

If you use llama.cpp through a language binding or through another application, you get these changes when that project moves to 1.0, not before. The launch post names no downstream projects and gives them no timeline, so check the release notes of the tool you actually install. For what running a large open model on your own hardware really demands in memory, yesterday's post on two open-weight releases has the numbers.

What changes if you run llama-server

The server got a round of hardening in the three days before the tag. Several things that worked before now need a flag, and one has no flag at all.

Before 1.0	In 1.0
Tools or MCP enabled on a non-loopback --host with no API key	Refused with an error. Set an API key, or opt in with --allow-unauthenticated-tools
A chat request with a media URL on your LAN, such as a file on a NAS	Blocked. Opt in with --media-allow-private-urls
The web UI's CORS proxy reaching an MCP server on localhost or the LAN	Not possible. The proxy only reaches public addresses, and the pull request says there is no escape hatch
--cors-origins with a specific origin	Now enforced properly, including the port. Foreign and null origins are rejected

The reasoning is in #32115: the server was designed on the assumption that users would sandbox it, but "many deployments are now run directly on local computer." A server on your laptop that will fetch any URL it is handed can be pointed at your router or your NAS. The companion pull request, #32113, describes its own changes as hardening and not as fixes for a known vulnerability, and admits they cost some convenience.

Under the new origin rules in #32098, same-origin requests, allowed origins and non-browser clients still get through. Scripts that call the API with curl or an SDK should not notice. A browser page on another origin will.

Three smaller server changes from the release notes:

• HF_TOKEN is only sent to huggingface.co hosts. A model URL on another host no longer receives your token.
• MCP tool calls are cancelled when the client disconnects or times out, and the MCP server is told.
• A new /v1/audio/speech endpoint follows the OpenAI API shape for text to speech.

If you use the ggml RPC backend to spread a model across machines, upgrade every machine together. The ggml 1.0.0 notes bump the RPC protocol to v9 and say incompatible peers are rejected.

What is new besides the cleanup

The release is not only removals. From the release notes:

• FP8 KV caches, selected with -ctk and -ctv f8_e4m3. The notes give no memory or quality figures, so none are repeated here.
• A host-memory cache for mixture-of-experts weights, with a configurable share kept on the CPU, exposed as moe_cache_size and moe_cache_cpu_share in llama_context_params.
• New model support: the Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0 text architectures, drama-3 text to speech, and MageVL-Cowork and Kimi-K3.1 vision.
• A redesigned web UI with chat groups, a live-preview markdown editor and a layout built for phones.

The launch post says the project supports more than 150 model architecture variants, three and a half years after it started.

One detail from the pruning pull request is worth a line. The project requires contributors to disclose AI use, and #30565 does: it lists a DeepSeek model in GGUF format, run through llama.cpp on a Mac Studio. The change that cleaned up the API for 1.0 was written with the help of a model running on the software being cleaned up. That is a smaller claim than the one AWS made earlier this week about its emulator, where no human reviewed the code. Here a named maintainer opened the pull request and it went through the project's normal review.

The short version

• llama.cpp 1.0.0 shipped on 13 November with ggml 1.0.0. Major versions now mean a breaking change to the public C API.
• If you call llama.h: 36 deprecated functions are gone, along with defrag_thold, --defrag-thold and clip.h. Run the search above, rename, and recompile.
• If you run llama-server: tools on a non-loopback host need an API key or --allow-unauthenticated-tools, private media URLs need --media-allow-private-urls, and the CORS proxy no longer reaches local addresses.
• If you use a binding or a wrapper: nothing changes until that project updates. Check its release notes.
• To get 1.0: Homebrew has it, or build the v1.0.0 tag with -DLLAMA_BUILD_IS_DEV=OFF. The numbered nightly builds continue alongside.

Frequently asked questions

Is llama.cpp 1.0 a breaking release?

Yes, for code that calls the C API. Version 1.0.0 removes 36 functions from llama.h that were already marked deprecated, drops the defrag_thold field and the --defrag-thold option, and removes clip.h. Each removed function except llama_set_warmup has a named replacement.

What replaces llama_load_model_from_file in llama.cpp 1.0?

llama_model_load_from_file. The related renames are llama_free_model to llama_model_free and llama_new_context_with_model to llama_init_from_model. The old names carried deprecation warnings before 1.0 and no longer exist in llama.h.

Does llama.cpp 1.0 change how llama-server behaves?

Yes. Tools or MCP on a non-loopback host without an API key are now refused unless you pass --allow-unauthenticated-tools, media URLs on private addresses need --media-allow-private-urls, and the web UI's CORS proxy can only reach public addresses.

How do I install llama.cpp 1.0.0?

Homebrew's llama.cpp formula listed 1.0.0 as stable on 14 November 2026. To build it yourself, check out the v1.0.0 tag and configure CMake with -DLLAMA_BUILD_IS_DEV=OFF. The GitHub release page has no binaries of its own and links to a nightly build instead.