Llama.cpp adds DeepSeek V4 support across platforms
DeepSeek V4 now runs on Llama.cpp with platform-specific optimizations...
The Llama.cpp team has enabled native support for DeepSeek V4 (MTP + DSpark variant) in their lightweight inference engine, addressing a major gap for developers seeking to run DeepSeek locally. This release (commit b10228) includes pre-optimized builds for over 20 hardware configurations, from consumer GPUs (NVIDIA CUDA 12/13, AMD ROCm) to mobile devices (iOS/Android ARM64). The GitHub repository now lists platform-specific binaries with verified signatures (GPG key B5690EEEBB952194) for security-conscious deployments.
Performance optimizations span Vulkan compute shaders, SYCL for Intel GPUs, and OpenVINO for Intel CPUs, with Apple platforms receiving special attention via KleidiAI acceleration for M-series chips. The update also introduces experimental support for openEuler's 310p/910b NPUs and Huawei's Ascend AI processors. Community reactions show strong interest, with 8 GitHub users reacting to the update within hours of its release.
For developers, this means DeepSeek V4 can now be deployed in air-gapped environments, on low-power devices, or with minimal latency for inference tasks requiring DeepSeek's advanced reasoning capabilities.
- DeepSeek V4 (MTP + DSpark) now natively supported in Llama.cpp across 20+ hardware platforms
- Pre-built binaries available for CPU/GPU (CUDA, Vulkan, ROCm) and mobile (iOS/Android) with verified GitHub signatures
- Optimized builds for Apple Silicon (KleidiAI), Intel GPUs (SYCL/OpenVINO), and experimental NPU support for openEuler/Huawei
Why It Matters
Enables offline, low-latency deployment of DeepSeek V4 across diverse hardware for privacy-focused or edge AI applications.