Ggml v1.0.0
Points and comments are a snapshot, not live.
GGML reaches v1.0.0 with deprecated API cleanup and broad backend performance improvements.
The release prunes deprecated API and legacy DIAG_MASK ops, adds configurable YaRN truncation and FP8 support. Backend work includes few-row quantized matmuls and MoE fusion on CUDA/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. The RPC protocol is bumped to v9 with faster loading and graceful failure handling.
What commenters are saying
Commenters largely celebrate the milestone, calling ggml and llama.cpp the "Linux of the AI world" for driving local inference across diverse hardware. Several users report consistent 5-10% speed improvements per upgrade, with one citing a runtime drop from 4:22 to 3:59 on a 50K-token prompt. A thread on naming notes that many projects (Linux, djgpp, dlmalloc) are named after their creators. Some note that specialized inference engines can be faster, but llama.cpp's value is its broad hardware and model support.