GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

llama.cpp Tags v0.1.0: First Semantic Version After Four Years of Daily Build Releases

The llama.cpp project tagged v0.1.0 on August 17, 2026 — the first release in the project’s history to carry a semantic version number. Since at least 2022, every release has been a build-numbered tag in the format b####. As of August 17, the most recent build tag was b10456.

The switch to semantic versioning is a signal about project maturity rather than a feature announcement. llama.cpp is the inference backend for Ollama, LM Studio, and dozens of other local AI tools. A stable versioning scheme makes dependency pinning and reproducible builds easier for the ecosystem built on top of it.

What Is in the Release

The v0.1.0 tag is concurrent with recent build releases. Material changes in the surrounding window:

  • Kimi-K3 text model support: Adds Moonshot AI’s 2.8T open-weight model (Kimi K3) to the supported architecture list, including chat template handling.
  • Server thread model redesign: A yield_to_queue thread model change intended to reduce latency under concurrent request load.
  • MTP assistant model loading: --models-dir now loads Multi-Token Prediction assistant models, enabling speculative decoding for models that ship MTP draft heads (including Qwen3.8-27B released three days earlier).
  • CUDA/SYCL kernel improvements: SYCL quantized copy kernel thread/block count now scales proportionally to the quantisation size. The Arc 70 q4_0 path goes from 20.21 GB/s to 158.19 GB/s throughput.
  • LoRA safety check: Validates LoRA tensor data is within file bounds before loading.
  • BoringSSL: Updated to 0.20260813.0.

The MTP and Kimi-K3 additions are the most relevant for current model usage. Qwen3.8-27B ships with MTP heads that llama.cpp can now use in draft mode (--spec-type draft-mtp), which independent testing puts at approximately 72% throughput improvement over single-token generation on the same hardware.

Why It Matters to the Ecosystem

llama.cpp is the lowest layer of the local inference stack. Ollama wraps it. LM Studio ships it. The GGUF format it defines is the packaging format for quantized open weights. When Alibaba releases Qwen3.8-27B community GGUF files on Hugging Face, those files run through llama.cpp.

The project has been actively maintained and heavily relied upon throughout that four-year period of build-numbered releases. The shift to semantic versioning is not a technical change — it is an operational one, signalling that the project intends to maintain a stable public API and that downstream tools can reason about compatibility in the standard way.

Build b10456 and v0.1.0 are the same codebase. The difference is the label. The label now means something.