The ggml-org project published ggml 1.0.0 on GitHub on Nov. 13, 2026, tagging commit 417a024 in a github-actions release signed on GitHub.com¹.

The project said the 1.0.0 cut prunes the deprecated public API and the legacy DIAG_MASK ops, adds configurable YaRN truncation and FP8 support, and includes performance work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends¹.

API notes in the release include removing DIAG_MASK_INF and DIAG_MASK_ZERO, bumping the RPC protocol so incompatible peers are rejected, and adding ggml_rope_set_truncate and ggml_rope_get_truncate to control YaRN correction-bound rounding¹.

The release also lists a backward pass for TRI, MUL_MAT src1 precision tests and gating for ggml_prec_set_acc, and an RPC protocol bump to v9 with faster loading and failure handling on unreachable endpoints or failed handshakes¹.

On the core side, the scheduler skips uploading inputs that are not read and prefetches only inputs produced by earlier splits, and an MoE cache can compute a share of cache misses on the CPU and copy results back asynchronously¹.

CUDA changes include few-row MMQ for batch sizes 2-24, MMVF for BF16 few-row matmuls, OCP FP8 support, fused NVFP4 gate/up matmuls with GLU in MMQ, and MoE layout and FlashAttention extensions¹.

Metal fused MoE routing and kept fusable node groups in a single command buffer; Vulkan added an Intel FlashAttention prefill kernel; SYCL added a device-side MoE scheduler; OpenCL added decomposed Adreno FlashAttention¹.

WebGPU gained batched tensor API support and new kernels including GROUP_NORM and MEAN; Hexagon added Q5_0/Q5_1 support and conv_3d; CPU paths switched MUL_MAT to F32 src1 and accumulate FlashAttention V in F32¹.

The release notes point to a ggml-org discussion on releases and versioning and ship source zip and tar.gz assets only¹.

Independent coverage of the tag was not in this assignment; downstream llama.cpp users would need to check compatibility after the API and RPC protocol breaks¹.