llama.cpp releases v1.0.0, ties major version bumps to API breaks

Hugging Face has released llama.cpp v1.0.0, introducing stricter semantic versioning rules and significant updates to API, model support, security, and UI. The release aims to provide a more stable and predictable development environment for downstream applications.

Orchestrated by Mira Sethi

November 13, 2026 at 9:34 AM EST2 min read

Key takeaways (3)

1.Under the new rule, MAJOR versions mark breaking changes to the public C API; the post says API and ABI breaks are covered.
2.Release notes say v1.0.0 removes 35 deprecated functions plus llama_set_warmup and defrag_thold, and renames llama_set_embd_state_src.
3.Release notes list llama-server fixes for SSRF, CORS and DNS rebinding but give no severity details or advisory.

Key takeaways (3)

1.Under the new rule, MAJOR versions mark breaking changes to the public C API; the post says API and ABI breaks are covered.
2.Release notes say v1.0.0 removes 35 deprecated functions plus llama_set_warmup and defrag_thold, and renames llama_set_embd_state_src.
3.Release notes list llama-server fixes for SSRF, CORS and DNS rebinding but give no severity details or advisory.

The llama.cpp team at Hugging Face released version 1.0.0 1 of the inference library and ggml on 13 November 2026, saying major-version bumps will now mark API and ABI breaking changes.

Developers who build on libllama or libggml can read the version number to judge whether an upgrade is likely to break their code. The release itself removes 35 deprecated functions and renames another, according to the v1.0.0 release notes 2.

Why the project adopted the rule

In its announcement post 1, the team said that in the 3.5 years since the project began, "many downstream applications already depend on our work." It said each new integration means more developers rely on its library interfaces, so "having a flexible but stable API is paramount."

The team wrote that it will be "stricter about making those boundaries visible to downstream consumers."

What the version rule says

According to the post, the rules apply to all ggml-org projects that use semantic versioning:

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

The post says both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps. It lists llama.cpp v1.0.0 (libllama) and ggml v1.0.0 (libggml) as the two releases.

API removals and additions

The release notes list these public header changes:

llama.h removes llama_set_warmup, 35 deprecated functions, defrag_thold and the --defrag-thold option.
The NextN API moves from llama-ext.h into llama.h, and llama_set_embd_state_src is renamed llama_set_ctx_state_src.
llama_context_params gains moe_cache_size and moe_cache_cpu_share.
clip.h is removed and its infrastructure consolidated into mtmd.
ggml 1.0.0 prunes its deprecated public API and the legacy DIAG_MASK ops.

The notes also list OCP FP8 KV caches (-ctk/-ctv f8_e4m3), a host-memory cache for Mixture of Experts (MoE) weights with a configurable CPU share, and loading of external PLE shards from separate GGUF files.

Server, models and interface

The release notes list llama-server fixes 2 for SSRF, CORS, DNS rebinding and MCP tool-call cancellation. They add two opt-in flags, --allow-unauthenticated-tools and --media-allow-private-urls, and restrict the HF_TOKEN to huggingface.co hosts. A separate change cancels MCP tool calls on client disconnect or timeout.

The notes do not describe the severity or impact of these items and do not reference a security advisory. The server also gains an OpenAI-compatible /v1/audio/speech endpoint.

The release notes 2 add three text architectures:

Kimi-K3.1-Pro, described as K3 with its full-attention MLA layers replaced by DSA.
Ripple-65B-Preview, described as a new dense 65B architecture.
xing4_0, described as combining mHC, MLA, MoE and MTP.

They also add drama-3 text-to-speech and vision support for MageVL-Cowork and Kimi-K3.1. The web UI is redesigned with named chat groups, a markdown editor and a mobile layout.

The post says llama.cpp currently supports more than 150 architecture variants. It says hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel is for self-hosted continuous-integration runners across major hardware platforms.

The security fixes and the stability goal are the project's own statements, and the post and release notes include no benchmark or independent test results. The post lists pre-release artifacts for Linux, macOS, Windows, Android and iOS, and conda-forge packages for Windows, macOS and Linux.

Microsoft finds editing fewer than 10 neurons can backdoor AI agents

Microsoft Research demonstrated that implanting a backdoor in an AI model requires editing fewer than 10 neurons, with the model emitting an attacker command only when a hidden trigger appears. The company said it has not observed the technique in the wild.

Adrian Cole · about 16 hours ago

UC Riverside models splitting HBM4 controller across host and base die

A UC Riverside paper evaluates four placements of HBM4 controller functions between host and base die, finding that moving maintenance to the base die frees host resources and that a die-to-die link saves host area and shoreline but shifts area and power to the base die.

Mira Sethi · about 18 hours ago

Gemini 4 Carbon tops MathArena leaderboard as Google releases Hydrogen

Gemini 4 Carbon reached first place on MathArena's competition-mathematics leaderboard with an 85.4% expected performance score, while Google's concurrent release of the faster Gemini 4 Hydrogen model extended the Gemini 4 family to consumer and developer tiers globally.

Elena Voss · about 21 hours ago