Open Source
Open Source

llama.cpp and ggml Ship v1.0.0: Deprecated APIs Removed

Reads only what the GitHub API fields, the two v1.0.0 release notes, Discussion 1579, LICENSE and README of the ggml-org llama.cpp and ggml repositories can verify. Both v1.0.0 tags landed on November 13, 2026. It separates semantic versioning, the rule that a release tag requires the internal ggml copy to match the released one, the split between b-numbered nightlies and vX.Y.Z releases, and the 35 removed functions, RPC v9 and MIT notice-retention and AS IS terms. The August 2026 versioning discussion is not presented as a November announcement, and no figure from the unretrieved HF blog is quoted.

Published November 15, 20269 min read
<!-- llama-cpp-ggml-1-0-0-opensource-resource | open-source | llama.cpp and ggml Ship v1.0.0: Deprecated APIs Removed -->

A project that had long tagged its nightly builds with a b plus a commit number suddenly gave both itself and its underlying tensor library a 1.0.0 on the same day, and cleared out a backlog of deprecated API in one pass. That gap between the two habits is more interesting than the version number alone. As of November 14, 2026, both ggml and llama.cpp under the ggml-org organization sit with v1.0.0 at the top of their tag lists, and the two v1.0.0 release notes both carry a published timestamp on November 13, 2026. This article stays with what the repository fields, the two release notes, the versioning discussion, LICENSE, and README can actually confirm; it does not supply benchmarks, user counts, or blog conclusions on the project's behalf. The repository path is ggml-org/llama.cpp, and the versioning rules trace back to https://github.com/ggml-org/ggml/discussions/1579.

A formal release, not another nightly

First, this is an official release rather than one more rolling nightly. The published_at values returned by the GitHub API are two UTC timestamps precise to the second: ggml v1.0.0 was published at 2026-11-13T08:34:21Z, and llama.cpp v1.0.0 at 2026-11-13T10:20:19Z. Both fall on November 13, 2026, with ggml first and llama.cpp after it. The top entry in each repository's tag list is v1.0.0, and the release html_url points to the respective /releases/tag/v1.0.0 page. In other words, 1.0.0 is a mark that can be checked on a release page, not a number that rolls forward every day in continuous integration.

The license of both repositories can be read directly from the /license endpoint: spdx_id is MIT, and the copyright line reads Copyright (c) 2023-2026 The ggml authors. The grant says that anyone obtaining a copy may, free of charge, use, copy, modify, merge, publish, distribute, sublicense, and sell copies. It also states one condition that must be met, namely that the copyright notice and the permission notice be included in all copies or substantial portions of the software, and it states word for word that the software is provided AS IS, without warranty of any kind. The accurate reading is therefore that MIT grants a wide freedom to use and redistribute while attaching one clear notice-retention condition and one disclaimer; it is not a grant with no conditions at all.

Why both repositories move on the same day

Tying the two repositories to a single release date is not a scheduling coincidence but a precondition written into the versioning policy. The original sentence in Discussion 1579 says that release tags are created on the condition that the internal ggml copy is identical to the released ggml copy, so that a downstream building from the release tags and linking to a system-wide ggml stays compatible. llama.cpp carries an embedded copy of ggml for its routine builds; if only one repository received the official mark while the other copy differed, downstream users who combined the library with a system-wide ggml could end up with symbols that do not line up. ggml therefore publishes its own v1.0.0 first, and llama.cpp follows by syncing the embedded copy and issuing a v1.0.0 with the same number.

The same discussion also spells out the split between two version-number schemes: daily nightly builds use a b plus a commit number as the Git tag, while official releases use semantic versioning of the form vX.Y.Z. That means b[NUM] and vX.Y.Z are not an old and a new spelling of one scheme but two parallel conventions serving rolling builds and public releases respectively, and 1.0.0 belongs to the second. The relationship can be reduced to two lines:

text
nightly : b[NUM]
release : vX.Y.Z

The discussion also names a header, llama-ext.h, whose purpose is to experiment with API changes without affecting the official API of the library and therefore without affecting the minor and major versions. A change can sit in that experimental layer until it is stable and only then be considered for promotion into the official public header. This is exactly the precondition that let the batch of API moves described below happen inside a single major bump: the old interfaces to be cleared had already been sorted between the experimental layer and the official one.

Where the semantic versioning rule comes from

One point of timing has to be exact, or the character of the whole article gets misread. The practice of semantic versioning, and the rule that breaking changes in ABI trigger a major release, come from Discussion 1579 opened by ggerganov in the ggml repository on August 7, 2026. Its body carries TODO items and community deliberation, so it is a draft-style policy discussion, not an announcement newly issued on November 13, 2026 alongside the two release notes. The More info section of the release notes merely lists that discussion link as further reading. What can be written is that the project already practices semantic versioning and treats ABI breakage as the trigger for a major bump; what must not be written is that this 1.0.0 announced a future policy of using major numbers to mark API and ABI incompatibility.

The decision rules in the discussion can be checked line by line: official release builds use semantic versioning vX.Y.Z; breaking changes in ABI trigger a major release, and while new function calls are usually kept backward compatible, adding a new enum value is still counted as a breaking change; for the server API, only breaking changes to inference endpoints such as completions and embeddings count as a major bump, while changes to the router API, tools, and CORS proxy are patches. Read against the two release notes, that explains why the version is 1.0.0 rather than one more increment on a 0.x line: this release concentrated on removals rather than additions alone. ABI is not a loose word here. It points to the binary contract a caller relies on when it compiles against a header and links against symbols, so once a public function is removed or renamed, an older program no longer runs unchanged on the new library.

What the major bump actually removed

The API changes section of llama.cpp is blunt: llama.h removes llama_set_warmup, 35 deprecated functions, defrag_thold, and the matching --defrag-thold flag, and the same batch promotes the NextN API from llama-ext.h into llama.h and renames llama_set_embd_state_src to llama_set_ctx_state_src. The ggml API changes say the same in kind, pruning the deprecated public API and removing the legacy DIAG_MASK_INF and DIAG_MASK_ZERO ops while bumping the RPC protocol to reject incompatible peers; a second line fixes the protocol number at RPC v9, with faster loading and graceful failure on unreachable endpoints or failed handshakes. Removals, a rename, a promotion, and a protocol bump together are the concrete content of "Deprecated APIs Removed", and they explain why this has to be an ABI watershed: a caller still compiling against the old headers or an old RPC peer is no longer silently tolerated.

One position change deserves its own note. The NextN and speculative API previously lived in the experimental header llama-ext.h, and this release lifts it into the official public header llama.h, which turns it from a changeable experimental interface into one that has to be honored under a version number. FP8 work is bound to the same major version: llama.cpp adds OCP FP8 KV caches, while ggml adds configurable YaRN truncation and FP8 support, and lists a host-memory MoE expert cache with a configurable CPU share plus loading of external PLE shards. The two release notes also list three new text architectures, Kimi-K3.1-Pro, Ripple and xing4_0, drama-3 TTS, and MageVL-Cowork and Kimi-K3.1 vision.

The server side is not cosmetic either. llama-server is hardened against SSRF, CORS, DNS rebinding, and MCP tool cancellation on client disconnect, adds a /v1/audio/speech endpoint, and offers explicit switches such as --allow-unauthenticated-tools and --media-allow-private-urls while restricting HF_TOKEN to the huggingface.co host. The direction of these switches is default deny with explicit opt-in: doors involving private addresses or unauthenticated tools stay shut unless the operator writes them open. Backend optimization in ggml is a separate thread, including few-row quantized matmuls and MoE fusion on CUDA and Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that avoid uploading unused inputs.

Seventeen backends and a compatibility surface that keeps getting patched

The README Supported backends table lists 17 backends word for word: BLAS, BLIS, CANN, CUDA, HIP, Hexagon, IBM zDNN, MUSA, Metal, OpenCL, OpenVINO, RPC, SYCL, VirtGPU, Vulkan, WebGPU, and ZenDNN, with OpenVINO marked In Progress. The description separately lists AVX, AVX2, AVX512, AMX, RVV and RISC-V, ARM NEON, 1.5 to 8 bit quantization, and CPU plus GPU hybrid inference. The table shows that a version bump has to account not for one device but for a cross-vendor hardware matrix; CANN maps to Huawei Ascend NPU, Hexagon to Snapdragon, IBM zDNN to IBM Z and LinuxONE, and Metal to Apple Silicon, and behind each cell sits a different compile and runtime path.

The many fix and workaround entries in the release notes point the same way: for example, Vulkan hiding internal symbols to prevent duplicate-dlopen state destruction, AMD proprietary drivers re-enabling coopmat after 26.10.1, and an 8-row limit on FlashAttention for large Qualcomm heads, all patches written against a specific driver or a specific op. With one pitfall per backend, a major bump landing on a single day rests on pulling the compatibility surface back to a boundary that can be verified, not on claiming that every backend is uniformly ready. The verifiable limits are written in the open as well: OpenVINO is still marked In Progress, and the GitHub API returns open_issues_count of 2597, which is a count of open issues, not a failure rate, and should not be read as a stability verdict. This article does not supply benchmarks, user counts, or blog conclusions for the project, and it does not quote any figure or conclusion from the HF blog llama-cpp-v1, which was not retrieved in this environment; what it writes is the part that can be traced word for word in repository fields, release notes, the discussion, LICENSE, and README.

The angle this article does not take

Several in-site articles already cover the path of putting local inference to work. Local LLM deployment comparison discusses deployment choices, Ollama local deployment SOP walks through setup steps, and VRAM deployment SOP works out VRAM budgets. The angle here is deliberately different: it is not about how to install a model or whether VRAM is enough, but about the semantics behind the v1.0.0 number, the ABI cleanup, and the two-repository sync mechanism. The same llama.cpp and ggml are an engine waiting to be deployed in those articles and an object being verified in this one, and neither substitutes for the other; readers after a hands-on path to a running service will find a dedicated in-site article more direct.

FAQ

Q1: Were llama.cpp and ggml 1.0.0 released on the same day? A1: Yes, both on November 13, 2026. The GitHub API published_at values are 2026-11-13T08:34:21Z for ggml and 2026-11-13T10:20:19Z for llama.cpp, v1.0.0 tops both tag lists, and this is an official release rather than a nightly.

Q2: Is 1.0.0 just a new number, or are there real breaking changes? A2: The release notes remove 35 deprecated functions, drop the DIAG_MASK ops, bump the RPC protocol to RPC v9 to reject incompatible peers, and rename llama_set_embd_state_src. These are ABI-level removals and incompatibilities, so a caller on old headers or an old peer is no longer silently tolerated.

Q3: Were semantic versioning and ABI-triggered major bumps announced on November 13? A3: No. Those rules come from Discussion 1579 opened on August 7, 2026, a draft-style discussion with TODO items and community deliberation; the November 13 release notes only list that discussion under More info, so it is not a same-day announcement.

Q4: Why do the two repositories have to ship together? A4: Discussion 1579 says a release tag is created only when llama.cpp's embedded ggml copy is identical to the released ggml, so downstream builds linking a system-wide ggml stay compatible. Nightly builds use a b plus commit number while official releases use vX.Y.Z, two schemes with different jobs, and llama-ext.h is the experimental layer that does not trigger a version bump.

Q5: Does MIT mean anything goes with no conditions attached? A5: No. Alongside the rights to use, copy, modify, distribute, and sell, the LICENSE text requires the copyright and permission notices to be kept in all copies or substantial portions, and states the software is provided AS IS without warranty. A wide grant and a disclaimer are not the same as no conditions, and treating them that way goes beyond the text.

Related reading

For inference frameworks and MoE running locally, see also layered inference SOP and MoE local inference open-source resource. Those two stay on the inference side and serve as background to the version semantics here; a dedicated in-site article is enough, and the two do not need to be merged into one.

This article is AI-assisted and human-edited. Last updated: 2026-11-15

FAQ

Were llama.cpp and ggml 1.0.0 released on the same day?
Yes, both on November 13, 2026. The GitHub API published_at values are 2026-11-13T08:34:21Z for ggml and 2026-11-13T10:20:19Z for llama.cpp, v1.0.0 tops both tag lists, and this is an official release rather than a nightly.
Is 1.0.0 just a new number, or are there real breaking changes?
The release notes remove 35 deprecated functions, drop the DIAG_MASK ops, bump the RPC protocol to RPC v9 to reject incompatible peers, and rename llama_set_embd_state_src. These are ABI-level removals and incompatibilities, so a caller on old headers or an old peer is no longer silently tolerated.
Were semantic versioning and ABI-triggered major bumps announced on November 13?
No. Those rules come from Discussion 1579 opened on August 7, 2026, a draft-style discussion with TODO items and community deliberation; the November 13 release notes only list that discussion under More info, so it is not a same-day announcement.
Why do the two repositories have to ship together?
Discussion 1579 says a release tag is created only when llama.cpp's embedded ggml copy is identical to the released ggml, so downstream builds linking a system-wide ggml stay compatible. Nightly builds use a b plus commit number while official releases use vX.Y.Z, two schemes with different jobs, and llama-ext.h is the experimental layer that does not trigger a version bump.
Does MIT mean anything goes with no conditions attached?
No. Alongside the rights to use, copy, modify, distribute, and sell, the LICENSE text requires the copyright and permission notices to be kept in all copies or substantial portions, and states the software is provided AS IS without warranty. A wide grant and a disclaimer are not the same as no conditions, and treating them that way goes beyond the text.

Related

Open Source

Univer Office SDK: Agents Edit Cells Without an AI SDK

Only what the dream-num/univer README, LICENSE and GitHub API fields support. Apache-2.0 covers the Web SDK core architecture and engine implementations; commercial plugins, the Server SDK and commercial AI SDK extensions are distributed under separate licences. Facade APIs such as FWorkbook, FWorksheet, FRange, getRange and setValues let an application or Agents Edit Cells directly with formula recalculation, while workbook.save() persists a JSON snapshot rather than XLSX import or export. No AI SDK is required anywhere in that path. Stars are cumulative GitHub API totals, not single-day growth.

Nov 16, 20269 min read
Open Source

LingBot-Touch: Code Apache-2.0, Weights Elsewhere

Robbyant/lingbot-touch reached 238 stars within days of creation, with Apache-2.0 in both the license field and the badge alt text, yet the model weights are not in the source repository: the README states verbatim that model files are hosted separately and access requires an account with permission to the model repository. This piece reads only API fields, README wording, and the LICENSE text: created November 9, 2026, last push November 11, main language Python, default branch main, not a fork and not archived. It also covers how three heterogeneous tactile sensors feed one representation and how the two model variants differ, without importing scores, download counts, or star totals the repository never states. As of November 12, 2026, code licensing and gated weights remain two separate questions.

Nov 13, 20269 min read