Skip to content

llama.cpp b10224: f16 repeat for WebGPU, CUDA 13 prebuilts

llama.cpp b10224 ships f16 repeat for the ggml-webgpu backend and prebuilt CUDA 13 DLLs across 30+ targets including ROCm 7.2, SYCL, and OpenCL Adreno.

news · llama-cpp · local-llm · release-notes
· 2 min

llama.cpp b10224 adds f16 repeat support to the ggml-webgpu backend — the single code change in this release.

PR #26307 brings half-precision repeat operations to WebGPU, which powers the browser-based inference playground at llama.app. The rest of the release is a prebuilt binary matrix covering 30+ targets.

Build targets in b10224:

  • macOS: Apple Silicon (arm64) with a KleidiAI-enabled variant. Intel (x64) marked DISABLED.
  • Linux: Ubuntu x64, arm64, s390x (CPU), plus Vulkan, ROCm 7.2, OpenVINO, SYCL FP32, SYCL FP16.
  • Windows: CPU, Vulkan, OpenVINO, SYCL, HIP, CUDA 12 (12.4 DLLs), CUDA 13 (13.3 DLLs), OpenCL Adreno.
  • Mobile: iOS XCFramework, Android arm64 (CPU).
  • openEuler: General build DISABLED. Ascend 310p and 910b (ACL Graph) variants still ship for x86 and aarch64.

CUDA 13 prebuilt DLLs are the quiet headline for local inference users. NVIDIA’s CUDA 13 toolkit is recent, and llama.cpp now ships 13.3-compatible binaries — no source compilation needed to run against the latest runtime.

NOTE

macOS Intel (x64) and the general openEuler build are both DISABLED in b10224. If you depend on either target, hold on b10223 until the next cycle.

Prebuilt binaries available now across all listed platforms. WebGPU backend with f16 repeat is live at llama.app.

Read next

Aug 8llama.cpp b10328: Docker tool isolation for local LLM serversAug 8llama.cpp b10331: server get_info reports correct working directory for tool isolatesAug 7llama.cpp b10306: SYCL GLU fast path, +14% f16 SWIGLU on Arc Pro B70