llama.cpp reaches 1.0 with new local AI inference features and explicit API versioning

What shipped on November 13

On November 13, 2026, ggml-org released llama.cpp v1.0.0 and ggml v1.0.0, then published a team announcement on Hugging Face. llama.cpp runs language and multimodal models through a portable C/C++ inference stack; ggml is its underlying machine-learning tensor engine. The official GitHub records put the llama.cpp release at 10:20 UTC and ggml at 08:34 UTC that day.

The releases are available now through their official GitHub tags. The llama.cpp record links nightly build b12359 under its assets guidance. The announcement lists distribution channels including Docker, Homebrew, conda-forge, Winget and XCFramework, but that list is not confirmation that every package channel already contains this version. Check the installed package and binary version rather than assuming a package-manager update resolves to 1.0.0.

There is no new hosted model/API identifier or metered API price in this announcement. v1.0.0 identifies the software release, not model weights or a cloud-service entitlement. Running it locally still requires a suitable model, hardware and compliance with that model’s own license.

The compatibility promise changes the upgrade decision

The team’s stated reason for the milestone is the growing number of downstream applications depending on its library interfaces. It says llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps. In its versioning table, a breaking public C API change under ./include gets a major bump; backward-compatible features, backend updates, model support and API additions get minor bumps; bug fixes without API changes get patch bumps.

That is useful information for teams embedding local inference in desktop apps, appliances or private services. It makes the intended compatibility boundary more visible. It does not mean old binaries become compatible with this release or that every later minor update will perform identically on every backend.

This release itself removes deprecated API surface. llama.cpp’s notes list removal of llama_set_warmup, 35 deprecated functions and defrag_thold, and promotion of the NextN/speculative API into llama.h. ggml’s notes also describe pruning its deprecated API and moving its RPC protocol to v9. Native bindings and distributed inference components therefore deserve a coordinated upgrade, not just a library-file swap.

New inference options and model coverage

llama.cpp’s release notes add OCP FP8 KV caches using -ctk/-ctv f8_e4m3, with optional static per-tensor scales. They also describe a host-memory mixture-of-experts cache with a configurable CPU share for cache misses, and loading external PLE shards. These are new configuration options, not evidence that a particular model will meet your memory, quality or latency budget.

The listed new text architectures include Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0. The notes also include drama-3 text-to-speech, MageVL-Cowork and Kimi-K3.1 vision support, plus Gemma 4 streaming audio input. Architecture support does not establish that every checkpoint, quantization or hardware backend works equally well; it also does not reduce a large model’s hardware requirements to those of a small local model.

The companion ggml release describes backend work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon. Examples include CUDA few-row quantized matrix multiplication, Metal MoE routing fusion and a Vulkan prefill FlashAttention path. These are maintainer-reported implementation changes. AccessAllGPT has not measured their throughput, memory impact or output quality, and no cross-hardware speedup is established by this briefing.

The local server gains speech and a refreshed interface

For application builders, llama-server adds an OpenAI-compatible /v1/audio/speech endpoint and a temperature option for /tts. That gives clients another familiar integration path, but endpoint compatibility should not be read as identical voice options, billing, behavior or full API parity with a hosted provider. The selected model must support the requested audio task.

The release also lists chat groups, a live-preview markdown editor and mobile layout changes in the web interface. Server changes include fixes around CORS, DNS rebinding and media URL handling, with explicit opt-ins for unauthenticated tools and private media URLs. These are release-note claims, not a security certification or a reason to expose a local inference server to the public internet without separate controls.

What local AI builders should do next

Our analysis: the most consequential change is the combination of a documented compatibility policy and a concrete migration event. Local AI products need repeatable releases as much as new model support. The policy helps teams plan dependency updates; the removed APIs make it important to test those plans now.

If your product embeds libllama or libggml, rebuild its bindings against the tagged headers and run your existing integration tests before rollout. If you use llama-server, replay a small, non-confidential set of real application requests against the old and new versions with the same model and hardware. Compare correctness, structured outputs, latency and peak memory, and exercise speech only if your application uses it.

Keep a known-working binary and configuration for rollback. Adopt when the exact workload passes, not because 1.0 implies universal production readiness. Teams without a current need for the new features can stage the upgrade while their bindings and packaging catch up.

Disclosures

AccessAllGPT received no payment, credits or private access for this briefing. ggml-org and Hugging Face did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent and is not affiliated with ggml-org or Hugging Face. Publication-wide relationships are listed on the disclosures page.