llama.cpp b10231 adds DSpark speculative decoding sidecars — and dspark now outranks dflash when no sidecar type is explicitly requested.
The dspark- files follow the same resolution rules as existing speculative sidecars:
-hfdtag applies to dspark files- Requested sidecar resolves without a full model at the tag
- Explicit
-mdselection disables sidecar discovery entirely - Auto-selection priority: dspark beats dflash because its sidecar carries an extra Markov head
NOTE
Speculative sidecars pair a small draft model with the main model to predict tokens ahead, cutting inference latency. DSpark’s sidecar carries an extra Markov head — the reason it outranks dflash in auto-selection.
Prebuilt binaries cover the usual matrix:
| Platform | Variants |
|---|---|
| macOS | arm64, x64 (KleidiAI disabled) |
| Linux | CPU (x64, arm64, s390x), Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16 |
| Windows | CPU, CUDA 12 (12.4 DLLs), CUDA 13 (13.3 DLLs), Vulkan, OpenVINO, SYCL, HIP, OpenCL Adreno |
| Mobile | iOS XCFramework, Android arm64 |
Windows x64 gets both CUDA 12.4 and CUDA 13.3 DLL builds — pick the one matching your toolkit.
Grab the release from GitHub. If you’re already using dflash sidecars, dspark takes over auto-selection unless you pass an explicit -md flag.