llama.cpp Releases

Added speculative decoding support (NextN/MTP) for GLM-5.2 (GLM_DSA) and updated the conversion tool to support `--mtp/--no-mtp` export options for this model.

Jul 29, 2026

Added support for the 'Laguna-S-2.1' Large Language Model type to the library.

Jul 28, 2026

Fixed critical WebGPU binding alias issues that prevented support on all architectures and resolved a recurrent-state-rollback test failure, alongside fixes for large buffer alignment and view offsets.

Jul 28, 2026

Fixed a critical bug in OpenCL Adreno image kernels that ignored the 4th dimension (ne3) of tensors, causing multi-stream batches to read incorrect data for key/value caches.

Jul 28, 2026

Refactored memory management calls by abstracting `llama_memory` to a common `common_memory` function to improve code structure and potential reuse.

Jul 28, 2026

This release focuses on internal graph optimizations in the ggml library, specifically addressing how view sources and outputs are handled to improve graph correctness and sampler stability.

Jul 28, 2026

Restores support for IQ4_NL quantization format in the Vulkan backend's Fully Attention (FA) mode to fix compatibility issues.

Jul 28, 2026

Accelerates Mamba-2 prefilling operations using a chunked SSD (Sliding Window) matrix multiplication implementation in CUDA.

Jul 28, 2026

Adds a Next.js B2C storefront to the Mercur marketplace CLI, making it a complete end-to-end solution with a live demo and various fixes.

Jul 28, 2026

Implements a Fast Walsh-Hadamard Transform (FWHT) kernel for the Metal backend on Apple Silicon to optimize performance.

Jul 28, 2026

Adds support for the Eagle3-V3 specification used in GPT-OSS models, enabling better compatibility with these architectures.

Jul 28, 2026

Disables the -ffast-math compiler flag for HIP (Radeon) support to improve numerical precision in calculations.

Jul 28, 2026

Adds support for the MiMo-V2.5 audio model, which utilizes Residual Vector Quantization (RVQ) for audio input processing.

Jul 27, 2026

Introduces a new utility function to print available computing devices, aiding in device detection and configuration.

Jul 27, 2026

Adds support for the Nanbeige4.2 model architecture, expanding the range of supported models in the library.

Jul 27, 2026

Optimizes GPU layer allocation for Multi-Threaded Parallel (MTP) blocks to ensure the front layers of the model remain on the GPU when using the n_gpu_layers parameter.

Jul 27, 2026

Parallelized the OpenCL compiler (ocloc) invocations in the SYCL build process for better build performance.

Jul 27, 2026

Adjusted the logic for offloading operations to the backend of the weights, and fixed graph issues for DeepSeek-V4.

Jul 27, 2026

Removed an unnecessary synchronization call in the state save/load test to improve performance.

Jul 27, 2026

Fixed a regression where explicitly specified draft models were silently overridden by automatic sidecar resolution, ensuring user configuration takes precedence.

Jul 27, 2026