llama.cpp b10021 boosts DeepSeekV4 with graph split reduction
New release optimizes inference for DeepSeekV4 and expands platform builds.
The llama.cpp project released version b10021, a maintenance update primarily aimed at optimizing the DeepSeekV4 model through reduced graph splits (PR #25702). This change streamlines the computational graph, leading to faster inference and lower memory usage when running DeepSeekV4 locally. The release also brings a comprehensive set of pre-built binaries across multiple platforms, including macOS (Apple Silicon with KleidiAI enabled, Intel x64, and iOS XCFramework), Linux (x64 and arm64 CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (x64 and arm64 CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android (arm64 CPU), and openEuler (x86 and aarch64 with ACL Graph). Additionally, UI assets have been updated. This release does not introduce new features but significantly improves the efficiency of DeepSeekV4 inference, making it more practical for resource-constrained environments.
- Reduced graph splits for DeepSeekV4 (PR #25702) improve inference speed and memory efficiency.
- Pre-built binaries for macOS (Apple Silicon with KleidiAI), Linux (ROCm, SYCL, Vulkan), Windows (CUDA 12/13, HIP), Android, and iOS.
- UI assets updated; release signed with verified GitHub GPG key B5690EEEBB952194.
Why It Matters
DeepSeekV4 optimizations make local LLM inference faster and more accessible across diverse hardware.