OpenToolslogo
ToolsExpertsNewsletter
AdvertiseLearn AI
  1. Home
  2. /
  3. News
  4. /
  5. llama.cpp and ggml reach v1.0.0 on 13 November 2026: what the semver rule covers, what actually broke, and where the 1.0 libraries ship
Updated 2 minutes ago

Share this article

PostShare

In This Article

  • A stability commitment stated for what comes next
  • The bump rules, in full
  • What 1.0.0 itself removed
  • The 1.0.0 tag also bundles additive work
  • Where the promise stops
  • Getting the 1.0 libraries today
  • What a pinning decision comes down to

AI news in your inbox

Weekly updates on tools, models, and the companies building them.

Subscribe

Footer

Company name

The right AI tool is out there. We'll help you find it.

LinkedInX

Knowledge Hub

  • News
  • Resources
  • Newsletter
  • Blog
  • AI Tool Reviews
  • YouTube Summary
  • YouTube Transcript Generator

Industry Hub

  • AI Companies
  • AI Tools
  • AI Models
  • MCP Servers
  • Muse Connectors
  • AI Tool Categories
  • Top AI Use Cases

For Builders

  • Submit a Tool
  • Submit a MCP
  • Experts & Agencies
  • Advertise
  • Compare Tools
  • Favourites

Legal

  • Privacy Policy
  • Terms of Service

© 2026 OpenTools - All rights reserved.

llama.cpp and ggml reach v1.0.0 on 13 November 2026: what the semver rule covers, what actually broke, and where the 1.0 libraries ship
Source: Official llama.cpp project mark. Source: ggml-org/llama-brand; reproduced under its commercial nominative-use permission. · Source image · CC BY-NC 4.0 with explicit additional commercial nominative-use permission in BRAND-USAGE.md

llama.cpp and ggml reach v1.0.0 on 13 November 2026: what the semver rule covers, what actually broke, and where the 1.0 libraries ship

By OpenTools Team · November 16, 2026

llama.cpp and ggml reached v1.0.0 on 13 November 2026, and the ggml‑org release post frames the milestone as a versioning commitment rather than a feature dump: from here on, both projects say they will express API and ABI incompatibilities through major‑version bumps. The rules are narrow — MAJOR means a breaking change to the public C API under ./include, MINOR covers backward‑compatible features, backend updates, model support or API additions, and PATCH is a bug fix with no API change. The versioned surfaces are libllama, with headers in llama.cpp/include/, and libggml, with headers in ggml/include/. 1.0.0 is itself the first breaking release under that rule. llama.h drops llama_set_warmup, the 35 functions marked deprecated, the defrag_thold field and the --defrag‑thold option, while the NextN/speculative API moves out of llama‑ext.h into llama.h and llama_set_embd_state_src becomes llama_set_ctx_state_src. ggml 1.0.0 prunes its deprecated public API, removes the legacy DIAG_MASK ops and bumps the RPC protocol to v9, which rejects incompatible peers. Two boundaries matter most for integrators. First, the promise as written covers the C headers only: clip.h was removed, mtmd headers changed, server endpoints and CLI flags changed, and stricter GGUF load requirements arrived — none of which a captured source places under the guarantee. Second, channel availability is not uniform. Homebrew's core formula reports stable 1.0.0 with prebuilt bottles, while conda‑forge still serves build‑numbered packages rather than the 1.0.0 tag. Docker, Debian, MacPorts and Winget were not verified, and post‑1.0.0 fixes so far exist only in nightly pre‑releases.

A stability commitment stated for what comes next

The Hugging Face release post, published 13 November 2026 by the ggml-org team, announces libraries v1.0.0 for both llama.cpp and ggml and frames the milestone as a commitment to downstream developers rather than a feature announcement. The post says that moving forward, both projects will reflect API and ABI incompatibilities with major-version bumps, and that the team will be stricter about making those boundaries visible as adoption grows.1

That promise is forward-looking. Nothing in the retained text claims the rule governed the 0.x line, so read it as the contract that begins at 1.0.0.

The versioned surfaces are two libraries with named public headers: libllama, for llama.cpp v1.0.0, with headers under llama.cpp/include/, and libggml, for ggml v1.0.0, with headers under ggml/include/.1 The post notes that many downstream applications already depend on these interfaces, which is the stated reason the boundary now needs to be visible.1

The bump rules, in full

The post states that the following rules apply to all ggml-org projects using semantic versioning:1

  • MAJOR — a breaking change to the public C API in ./include.
  • MINOR — a backward-compatible feature, backend updates, model support, or an API addition.
  • PATCH — a bug fix with no API change.

The scope is tighter than the headline commitment. The MAJOR row is written around the public C API in ./include; ABI appears in the prose sentence above the table, not inside the table row itself. The table also names only the two libraries above — mtmd and the other components this release touches are not listed there.1

What 1.0.0 itself removed

The tag that establishes the rule is also its first test case: 1.0.0 is a breaking release for both libraries.

For llama.cpp, the release notes list llama.h removing llama_set_warmup, 35 deprecated functions, the defrag_thold field and the --defrag-thold option, and describe the NextN/speculative API being promoted from llama-ext.h into llama.h with llama_set_embd_state_src renamed to llama_set_ctx_state_src.2 Multimodal consumers lose clip.h, removed and consolidated into mtmd in the same release.2

The change that defines that set is pull request #30565, titled llama : prune, merged on 12 November 2026 from six commits into commit 8ca5a80.3 Its description identifies the removals as llama_set_warmup, all functions marked DEPRECATED in llama.h, the deprecated llama_context_params::defrag_thold field and the -dt/--defrag-thold option, and describes promoting the NextN functions into llama.h alongside the rename.3 It does not itemize the 35 names in the retained text; they are identified by their DEPRECATED markers in pre-1.0.0 headers. Downstream impact is already visible: the same thread cross-references a matching removal and rename pull request in the hybridgroup/yzma project, dated 13 November 2026 and marked merged.3

ggml 1.0.0 is independently breaking. Its release notes say ggml prunes the deprecated public API, removes the legacy DIAG_MASK_INF and DIAG_MASK_ZERO ops, and bumps the RPC protocol to v9 so incompatible peers are rejected.4

Migration guidance is partial. A maintainer suggested in the thread that the release notes point readers to the deprecation notices in pre-1.0.0 versions for replacements; the retained release text does not show that note, so whether it shipped is unconfirmed.3 A separate migration guide for the llama_batch API exists as an open pull request, #32163.3 The older llama_batch_* functions were not removed in this bump: the thread records a proposal to deprecate them in favour of llama_batch_ext_*, a maintainer reply that removing them now is too early, and a suggestion that deprecation could follow after the v1 release.3

The 1.0.0 tag also bundles additive work

Reading 1.0.0 as if everything since 0.9.x were breaking would over-read the version number. The same release adds llama_context_params::moe_cache_size and moe_cache_cpu_share, new mtmd audio handling, and an OpenAI-compatible /v1/audio/speech endpoint on the server.2 Those are the kind of additions the MINOR rule describes; they shipped inside a MAJOR bump because of the removals above, not because of themselves.1

Where the promise stops

Several surfaces changed at 1.0.0 without a stated guarantee:

  • mtmd headers. The versioned-releases table lists libllama and libggml only, yet this release removes clip.h and changes mtmd.h and mtmd-helper.h 1.2 No retained source says whether mtmd carries the same promise.
  • The server and CLI. --defrag-thold was removed in this bump and server endpoints changed, but the rules table scopes a major bump to the public C API in ./include. Nothing captured states a compatibility rule for the HTTP API, CLI flags or the bundled web UI 1.2 One review comment notes that a router preset which still sets defrag-thold will now stop the router at startup, because the router refuses unknown keys.3
  • GGUF model files. 1.0.0 adds load-time requirements, including requiring the projector and patch-embedding tensors, a 2x2 spatial merge for some projectors, and an F32 position embedding for HunyuanVL.2 Those can affect older model files, and no captured text states a file-format compatibility policy.
  • Support duration. No source in this package commits to a maintenance window for the 1.x line.

One further limit: the itemized list of the 35 removed functions is not in the retained evidence, because it sits in the #30565 diff, which this capture does not include.3

Getting the 1.0 libraries today

The post names the channels and the platforms each one targets:1

  • Pre-release artifacts — Linux, macOS, Windows, Android, iOS
  • Docker — Linux
  • Debian — Linux
  • Homebrew — macOS, Linux
  • conda-forge — Windows, macOS, Linux
  • MacPorts — macOS
  • Winget — Windows
  • XCFramework — iOS, macOS

Checked after the announcement, two of those channels differ from each other:

  • Homebrew carries 1.0.0. The homebrew/core formula reports stable 1.0.0, pinned to tag v1.0.0 at revision b1b770d90d94cad9e9ea5b993bdbc383a522fb28, with prebuilt bottles for arm64 macOS (golden_gate, tahoe, sequoia), arm64_linux and x86_64_linux.5 It depends on a separate ggml formula plus openssl@4, and that formula's version was not captured — so this record does not establish which libggml Homebrew delivers.5
  • conda-forge does not. Its llama.cpp package reports latest_version 12418 across linux-64, linux-aarch64, win-64, osx-64 and osx-arm64, last modified 15 November 2026 — a build-number style version rather than the 1.0.0 tag.6 Reading 12418 as a specific nightly build is inference from naming convention, not a captured statement.

Docker, Debian, MacPorts and Winget were not verified in this package, so 1.0.0 availability there is unknown rather than absent. The blog also links a discussion for packagers on adding ggml-org projects to packaging and release pipelines, without showing package names.1

v1.0.0 remains the latest stable tag in both repositories, and post-tag work lands in nightly pre-releases; the llama.cpp releases page listed builds b12427 through b12441 for 15–16 November 2026 7.8 Those nightlies carry fixes for problems observed after the tag, including a CUDA MUL_MAT_ID abort with odd src0 columns on AMD hardware (b12433) and an MoE regression on HIP (b12430).7 Whether the underlying issues also affect the v1.0.0 tag is not established by the retained text, so treat that as a possibility to check rather than a fact.

What a pinning decision comes down to

On this evidence the practical reading is narrow. The only surface the 1.0.0 rule protects is the public C API in llama.cpp/include/ and ggml/include/, with ABI incompatibility described in the surrounding prose rather than in the table.1 Everything else that moved in this release — clip.h removal, mtmd.h and mtmd-helper.h changes, the --defrag-thold removal, and stricter GGUF load requirements — sits outside a stated guarantee, so an upgrade plan should not assume those surfaces are covered.2

For shipping, verify the channel directly rather than trusting the channel list: one checked source reports Homebrew at 1.0.0, another reports conda-forge still on build-numbered packages, and the rest were not verified here 5.6 If a project needs fixes that currently exist only in the post-1.0.0 nightlies, that is a deliberate move off the stable tag, because pre-releases are not the versioned release the promise is written around.7

Sources

  1. 1.huggingface.co (huggingface.co)
  2. 2.github.com (github.com)
  3. 3.github.com (github.com)
  4. 4.github.com (github.com)
  5. 5.formulae.brew.sh (formulae.brew.sh)
  6. 6.api.anaconda.org (api.anaconda.org)
  7. 7.github.com (github.com)
  8. 8.github.com (github.com)