기본 콘텐츠로 건너뛰기

llama.cpp 1.0 Upgrade: What Breaks and How to Fix It

llama.cpp v1.0.0, released on November 13, 2026, is the first version where the project promises semantic versioning for its C library: from now on a breaking change to the public API in ./include means a major-version bump. To get there, the release deletes the old deprecated functions, a KV-cache option and the clip.h header, and tightens several llama-server security defaults. This guide lists what can break when you upgrade from v0.9.x, how to fix each case, and what is worth trying once you are on 1.0.

Who actually needs to worry

Not everyone who runs llama.cpp hits a breaking change. Before you read the whole list, place yourself in one of three groups:

  • You only run the prebuilt binaries (llama-cli, llama-server) with a model and a few flags. Most of your setup keeps working. Check two things: whether your scripts pass -dt / --defrag-thold, and whether your server uses tools, MCP, a CORS proxy or media URLs on a local network.
  • You link against libllama from C, C++ or a language binding. This is the group the release is about. If your code still calls functions that were marked DEPRECATED in llama.h, it will no longer compile, because those declarations are gone.
  • You build on the multimodal code and include clip.h directly. That header was removed and its code was consolidated into mtmd.

If you use llama.cpp only through another app, the upgrade is that app's job: wait for its maintainers to ship a build against v1.0.0 rather than swapping the library underneath it.

What the 1.0 version number promises

The Hugging Face blog post that accompanied the release explains the versioning rules that now apply to ggml-org projects. A MAJOR bump means a breaking change to the public C API in ./include; a MINOR bump covers backward-compatible features, backend updates, model support or API additions; a PATCH bump is a bug fix with no API change. Two libraries are covered: libllama (public headers in llama.cpp/include/) and libggml (headers in ggml/include/), and ggml was bumped to v1.0.0 at the same time.

In practice this gives you something you did not have before: you can pin 1.x in a package manager or CMake script and expect your code to keep compiling across minor releases. Nightly build tags such as b12359 still exist next to the semantic versions, so decide which one your build pins and write it down.

Breaking change 1: deprecated llama.h functions are gone

Pull request #30565 ("llama : prune") removed all 35 functions that were marked DEPRECATED in llama.h, plus llama_set_warmup. The good news is that every removed declaration carried a message naming its replacement, so the migration is mostly search-and-replace. The ones you are most likely to have in older code:

  • 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 → the llama_model_n_* versions
  • llama_n_vocab → llama_vocab_n_tokens
  • llama_token_bos, _eos, _eot, _nl, _pad and the llama_token_fim_* family → the matching llama_vocab_* functions
  • llama_token_get_text / _score / _attr and llama_token_is_eog / _is_control → llama_vocab_get_* and llama_vocab_is_*
  • llama_add_bos_token / llama_add_eos_token → llama_vocab_get_add_bos / llama_vocab_get_add_eos
  • llama_get_state_size, llama_copy_state_data, llama_set_state_data → llama_state_get_size, llama_state_get_data, llama_state_set_data
  • llama_load_session_file / llama_save_session_file → llama_state_load_file / llama_state_save_file
  • llama_sampler_init_grammar_lazy → llama_sampler_init_grammar_lazy_patterns

Note that the token helpers now take a llama_vocab pointer, not the model. If your old code passed the model around, get the vocab once after loading and pass that.

How to fix it step by step

  1. Before upgrading, build your project against v0.9.x with deprecation warnings visible (do not silence -Wdeprecated-declarations). Every warning is a line that will fail on 1.0.
  2. Replace each call with the name from the deprecation message (the list above, or our table at the end of this post).
  3. Rebuild against v0.9.x until the warnings are gone. Your code is now compatible with both versions, which makes the switch safe to roll back.
  4. Bump to v1.0.0 and rebuild. Anything left is either llama_set_warmup, defrag_thold or the renamed state function covered below.

llama_set_warmup and the renamed state call

llama_set_warmup has no direct replacement; its deprecation note said user code should do warmup runs manually. If you relied on it, run a short throwaway decode yourself before serving real requests. Separately, the speculative-decoding ("NextN") functions that lived in llama-ext.h were promoted to llama.h, and llama_set_embd_state_src was renamed to llama_set_ctx_state_src. That one is a pure rename.

Breaking change 2: --defrag-thold and defrag_thold

The same pull request removed the deprecated defrag_thold field from llama_context_params and the -dt / --defrag-thold command-line option. If a systemd unit, Docker command or launcher script still passes the flag, remove it; the option no longer exists in v1.0.0's argument list. In C/C++ code, delete the line that sets params.defrag_thold.

Breaking change 3: clip.h was removed

Pull request #31865 removed clip.h and moved the vision/audio model infrastructure into mtmd, with mtmd_context as the public handle. The PR author describes clip.h as an internal header that no longer made sense on its own. If your code included it, move to the public mtmd.h / mtmd-helper.h APIs; there is no compatibility shim. The release also makes some vision projectors stricter at load time (required projector and patch-embedding tensors, a 2x2 spatial merge for Qwen-VL projectors), so a malformed or very old mmproj file that used to load may now be rejected. Re-download or re-convert it.

Breaking change 4: llama-server security defaults

This is the change most likely to surprise people who never touch C code. Several server behaviours that used to be allowed now need an explicit opt-in:

  • Tools or MCP on a network address without an API key are refused. If you start the server with --host set to a non-loopback address and enable tools, MCP servers or the MCP proxy without an API key, it is now refused with an error. Fix: set --api-key (or --api-key-file). The escape hatch --allow-unauthenticated-tools exists, and the docs describe it as "only for trusted networks".
  • Media URLs to private addresses are blocked. An image_url pointing at a NAS or another machine on your LAN is now refused by default. Fix: add --media-allow-private-urls if you trust that network, or send the image as base64 instead.
  • The CORS proxy only reaches public addresses. According to PR #32115, using the web UI's CORS proxy to reach an MCP server on localhost or your LAN is no longer possible, and there is no flag to bring it back. Point the client directly at the MCP server instead.
  • --cors-origins now checks the port. An origin written as http://domain:PORT must match exactly, so a frontend on a different port than the one you listed will start failing CORS checks. List every origin you actually use.
  • HF_TOKEN is only sent to huggingface.co hosts, so a custom mirror that relied on receiving it will now see anonymous requests.

Comparison: where llama.cpp 1.0 sits next to Ollama and vLLM

All three are free, open-source local inference engines; what differs is the interface and how they version. Facts below come from each project's GitHub repository and latest release, checked today.

Project License / price Latest release (as of Nov 15, 2026) Main interface Versioning note
llama.cpp MIT, free v1.0.0 (Nov 13, 2026) C/C++ library (libllama), CLI, llama-server with OpenAI-compatible endpoints Semantic versioning for the public C API starts with 1.0
Ollama MIT, free v0.40.16 (Nov 9, 2026) Go app with its own CLI and local API Still on 0.x releases
vLLM Apache-2.0, free v0.33.0rc3 (Nov 2, 2026) Python serving engine for GPU servers Still on 0.x releases

If you embed inference in your own program, llama.cpp is now the only one of the three with a written API-stability promise. If you just want a local chat model with minimal setup, the upgrade notes above matter less than whichever tool you already know.

Worth trying after you upgrade

  • /v1/audio/speech. An OpenAI-compatible text-to-speech endpoint for models that can generate audio. Read the details before you drop in an OpenAI client: the default format is wav (not mp3), only wav and pcm are accepted, OpenAI voice names like alloy are not mapped, speed must be 1.0, and the sample rate is the model's own rather than 24 kHz.
  • FP8 KV cache. -ctk f8_e4m3 -ctv f8_e4m3 stores the KV cache in FP8, with optional static per-tensor scales.
  • MoE expert cache. New moe_cache_size and moe_cache_cpu_share context parameters keep MoE experts in host memory with a configurable CPU share.
  • Redesigned web UI. Chat groups in the sidebar, a live-preview markdown editor, PDF thumbnails and a mobile layout.

Editor's note

Upgrade now if: you maintain code against libllama and want the stable-API guarantee, or you expose llama-server on a network and want the tighter defaults.

Wait if: you use llama.cpp only through another app, or you depend on reaching LAN MCP servers through the web UI's CORS proxy, which has no opt-in.

Common errors and fixes: "undeclared identifier llama_load_model_from_file" (or any old llama_token_* name) means a removed deprecated call: use the replacement above. An unknown-argument error for -dt means delete the flag. The server refusing to start with tools on 0.0.0.0 means you need --api-key. Images from a LAN URL failing means --media-allow-private-urls or base64.

What we checked: everything here comes from the v1.0.0 release notes, the linked pull requests, the deprecation messages in the removed llama.h lines and the v1.0.0 server README. We did not benchmark the new features, so this post makes no speed claims.

FAQ

Do I have to rewrite my app to use llama.cpp 1.0?

No. If you already moved off the deprecated functions, most code compiles unchanged. If not, the work is mostly renaming calls, because each removed function had a named replacement.

Will my GGUF models still work?

The release does not announce a GGUF format break. The exception is multimodal projector (mmproj) files, which now get stricter checks at load time; if one is rejected, get a fresh copy of it.

Can I stay on v0.9.0?

Yes, older releases remain on GitHub. But the security fixes for llama-server (CORS, DNS rebinding, SSRF) are in the 1.0 line, so do not leave an old server exposed on a network.

Related reading on lightmind24

Bottom line: fix the deprecation warnings on v0.9.x first, then upgrade, and check your server flags if it listens on anything other than localhost.

Screenshots of the llama.cpp v1.0.0 release notes and the versioning rules, plus our own migration table, are included at the end of this post. Image credit: screenshots from github.com/ggml-org/llama.cpp and huggingface.co/blog; migration table by lightmind24.

lightmind24 is an independent review site, not affiliated with ggml-org, Hugging Face, Ollama or vLLM.

Last verified: 2026-11-15

Sources:
llama.cpp v1.0.0 release notes
llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem (Hugging Face blog)
PR #30565: llama : prune
PR #31865: remove clip.h, consolidate into mtmd
PR #32113: explicit opt-in tools via non-loopback address
PR #32115: disallow local addresses from CORS proxy and media URLs
llama-server README (v1.0.0)
Ollama on GitHub
vLLM on GitHub

Uploaded Image Uploaded Image Uploaded Image Uploaded Image

댓글

이 블로그의 인기 게시물

브루(Vrew) 요금제 정리: 무료 200크레딧으로 어디까지 쓸 수 있나

요약: 영상 자막 편집기 브루(Vrew)는 2026년 4월 22일부터 기능별 사용량 대신 '통합 크레딧' 하나로 요금을 매겨요. 무료 플랜은 매달 200크레딧이 들어오는데, 음성 분석(자동 자막)만 쓰면 최대 200분 분량이라 짧은 영상 몇 편은 충분하지만, 영상 초반의 Vrew 마크를 지우려면 유료 플랜이 필요해요. 이 글은 브루 공식 가격 페이지와 고객센터 공지를 2026년 10월 23일에 다시 확인해, 크레딧 계산법과 플랜 고르는 기준을 정리했어요. 목차 브루 요금 구조: 통합 크레딧이란 1크레딧으로 할 수 있는 일 무료 플랜으로 되는 것과 안 되는 것 유료 플랜 4종 비교표 월 구독·연 구독·기간 이용권, 무엇이 이득일까 상황별 추천: 학생·직장인·유튜버 크레딧이 모자랄 때: 사용 기록 지우개 업그레이드·환불 전에 알아둘 규칙 운영자 메모, 자주 묻는 질문 1. 브루 요금 구조: 통합 크레딧이란 예전 브루는 플랜마다 음성 분석 몇 분, AI 목소리 몇 자, 번역 몇 자처럼 기능별 한도를 따로 줬어요. 그런데 보이저엑스(브루 운영사)가 실제 이용 데이터를 보니 한두 가지 기능만 집중해서 쓰는 사람이 훨씬 많았다고 해요. 그래서 2026년 4월 22일 업데이트부터는 플랜마다 정해진 크레딧을 받고, 어떤 기능에 쓸지는 내가 정하는 방식으로 바뀌었어요. 플랜별 매달 지급량은 Free 200 · Light 2,000 · Standard 10,000 · Business 12,000크레딧 이에요. 공식 공지에 따르면 무료 사용자는 매달 1일에, 유료 사용자는 매월 결제일에 충전되고, 남은 크레딧은 다음 달로 넘어가지 않아요. 바뀐 뒤 무엇이 좋아졌나 공지에 나온 예시가 이해하기 쉬워요. Standard 플랜에서 AI 목소리만 쓰면 예전에는 최대 50만 자였지만, 지금은 1만 크레딧을 모두 AI 목소리에 몰아 쓸 수 있어 최대 100만 자가 돼요. 반대로 여러 기능을 골고루 쓰는 사람은 크레딧이 기능끼리 ...

에이닷 전화 vs 익시오: 통화 녹음·AI 요약, 내 통신사·폰에 맞는 앱 고르기

요약: 통화 녹음과 AI 요약을 쓰고 싶다면 먼저 내 통신사와 휴대폰 종류부터 확인해야 해요. LG유플러스 고객은 무료 앱 익시오, SKT 고객은 에이닷 전화가 가장 기능이 많고, 다른 통신사의 아이폰 사용자는 iOS에 들어 있는 기본 통화 녹음·전사 기능이 현실적인 선택이에요. 이 글은 2026-10-30에 SKT·LG유플러스·Apple 공식 페이지와 도움말을 직접 확인해 세 가지를 비교했어요. 왜 '앱 기능'보다 '내 조건'을 먼저 봐야 할까요 통화 요약 앱 비교 글을 보면 기능 목록부터 나열하는 경우가 많아요. 그런데 실제로 써 보려는 사람에게 가장 먼저 걸리는 건 기능이 아니라 이용 조건 이에요. 익시오는 LG유플러스 고객 전용이고, 에이닷 전화는 누구나 설치할 수 있지만 통역콜·baro 같은 기능은 SKT 고객에게만 열려 있어요. 아이폰은 운영체제 구조상 앱이 통화를 마음대로 녹음하기 어려워서, 통신사 앱이 안 되면 iOS 기본 기능에 기대야 해요. 그래서 이 글은 "어떤 앱이 제일 좋은가"보다 "내 통신사·내 폰에서 무엇을 쓸 수 있나" 순서로 정리했어요. 기능 차이는 그다음이에요. 한눈에 비교: 에이닷 전화 vs 익시오 vs 아이폰 기본 통화 녹음 구분 에이닷 전화 (SKT) 익시오 ixi-O (LG유플러스) 아이폰 기본 통화 녹음 (Apple) 누가 쓸 수 있나 통신사 구분 없이 앱 설치 가능. 단, 통역콜·baro·그룹통화 발신 등은 SKT 고객 전용 U+ 모바일 고객. 타 통신사, 만 14세 미만, 선불요금제 고객은 제한 (공식 사이트 FAQ상 알뜰폰 중 KB리브엠은 예외) 지원 국가·언어의 iPhone 사용자. 1명과 통화할 때만 녹음 가능 지원 OS·기기 Android 10 이상, iOS 16 이상 iOS 17 이상 iPhone 12 이상 / Android...

타입캐스트 vs 클로바더빙: 수익 나는 채널에 쓸 때 무료·유료 조건 비교

요약 타입캐스트와 클로바더빙은 둘 다 글만 입력하면 한국어 AI 목소리를 만들어 주는 서비스지만, 무료로 쓸 수 있는 범위가 꽤 다릅니다. 클로바더빙 무료는 매달 15,000자까지 쓸 수 있는 대신 수익이 생기는 채널에는 쓸 수 없고, 타입캐스트 무료는 처음에 주는 약 5분 분량의 다운로드로 끝나는 체험판에 가깝습니다. 광고 수익이 붙은 유튜브나 쇼핑몰 영상에 넣을 목소리라면 처음부터 유료 플랜 조건을 보고 고르는 편이 안전합니다. 왜 이 두 서비스를 같이 비교하나 쇼츠 내레이션, 수업 자료 더빙, 상품 소개 영상처럼 "내 목소리 대신 AI 목소리"가 필요할 때 한국 사용자들이 가장 많이 찾는 곳이 네오사피엔스의 타입캐스트 와 네이버의 클로바더빙 입니다. 기능 소개만 보면 비슷해 보이지만, 실제로 문제가 되는 건 음질보다 "이 목소리를 내 채널에 써도 되나" 입니다. 그래서 이 글은 기능 나열보다 무료 사용 범위, 출처 표기 의무, 유료 플랜 가격과 다운로드 한도를 중심으로 정리했습니다. 모든 수치는 2026년 10월 25일에 두 서비스의 공식 요금 페이지와 고객센터 도움말에서 직접 다시 확인한 내용입니다. 한눈에 보는 비교표 항목 타입캐스트 무료 타입캐스트 베이직 클로바더빙 Free 클로바더빙 Standard 월 요금 0원 9,900원 (연간 결제 시 10% 할인) 0원 19,900원 (네이버 클라우드 플랫폼에서 결제) 무료·포함 사용량 음성 생성·재생 무제한, 다운로드는 최초 3,000크레딧(약 5분) 1회 지급 월 30,000크레딧(약 60분) 월 15,000자, 다운로드 월 20회, 프로젝트 누적 5개 (매월 1일 갱신) 월 30,000자, 다운로드 월 40회, 프로젝트 누적 20개 쓸 수 있...