llama.cpp Releases

Introduces unit tests for the `llama-batch` system and fixes Windows build issues.

Jul 10, 2026

Adds OpenCL cluster-parallel decode FlashAttention support for Adreno GPUs, optimizing performance on Android and Windows devices.

Jul 10, 2026

Optimized CUDA top-k and argsort implementations to reduce temporary buffer memory usage.

Jul 10, 2026

Added --output option to CLI for specifying output file path.

Jul 10, 2026

Added optimizations for Hexagon (Qualcomm) backend including tiling and fast division.

Jul 9, 2026

Fixed crash when using chat templates that fail at apply time but pass at init time.

Jul 9, 2026

Optimized compilation by avoiding reliance on compile-time optimizations and using cheaper arithmetic operations.

Jul 9, 2026

This release is NOT from ggml-org/llama.cpp. It is from mercurjs/mercur repository.

Jul 9, 2026

This release is NOT from ggml-org/llama.cpp. It is from mercurjs/mercur repository.

Jul 9, 2026

Added offline parameter support to the llama-bench benchmarking tool.

Jul 9, 2026

Added depthwise convolution support for Metal backend and fixed F16 kernel support for the new operation.

Jul 9, 2026

Enabled fast math optimizations for HIP/AMD GPU backend to match CUDA performance.

Jul 9, 2026

CUDA backend fix aligning snake fusion matcher logic with other backends.

Jul 9, 2026

Server-side fix ensuring prompt batch splitting respects the minimum step size.

Jul 9, 2026

Adds RoPE (Rotary Positional Embedding) support for Vision tasks in the Hexagon backend.

Jul 9, 2026

Tuning of subgroup splitting in WebGPU for better Flash Attention vector operations.

Jul 9, 2026

Fixes garbled output in OpenCL for Q6_K weights with specific alignment requirements.

Jul 8, 2026

Vulkan backend performance tuning for GCN GPUs and attention masking optimization.

Jul 8, 2026

OpenCL optimizations for MoE (Mixture of Experts) prefill operations.

Jul 8, 2026

Fixes a bug in the llama-batch system allowing positions to decrease within a sequence.

Jul 8, 2026