llama.cpp v1.0.0 Released with CUDA and Vulkan Optimizations
The llama.cpp project and ggml library have officially released major version v1.0.0 (build b12359). The release solidifies the cleanup of legacy code and a series of low-level optimizations for various hardware architectures.
Inference Optimizations
The CUDA backend adds few-row MMQ kernels for small batches sized 2-24. Prefill for NVFP4 format checkpoints is optimized by merging w_s scales in the MMQ epilogue. For Intel graphics, the Vulkan backend integrates a FlashAttention kernel during the prefill stage.
Processor and NPU Support
The following improvements have been implemented on the CPU and NPU side:
- Added 4x4 simd_gemm tiles for ARM NEON
- Implemented GATED_DELTA_NET routing to SpacemiT RVV vector cores (RISC-V)
- Added support for Q5_0/Q5_1 quantization and host build for Windows ARM64 for Qualcomm Hexagon
Server-Side Security
The llama-server implements cancellation of MCP tool calls upon client disconnect or timeout. SSRF vulnerabilities in the CORS proxy have been closed, and access to tools from non-local addresses has been tightened. Fixes include Qwen3.5 model conversion and multimodal projector issues.











