llama.cpp Releases

Optimized imatrix calculation by moving finite checks and limiting validation to only the experts that were touched during the run.

Aug 12, 2026

Added support for Pocket-TTS in Multi-Modal models, optimizing convolution operations to reduce generation latency significantly.

Aug 12, 2026

Updated tests to skip multi-output sampling on HIP due to missing CUB support, and improved CI logging for ROCm self-hosted runners.

Aug 11, 2026

Fixed a critical bug where Sliding Window Attention (SWA) was not being enabled for EXAONE 4.5 models due to incorrect parameter reading order.

Aug 11, 2026

Fixed the PEG parser to suppress warnings/errors associated with incomplete escape sequences in input parsing.

Aug 11, 2026

Fixed CI errors in the WebGPU backend and added i32 support to copy operations to enable comprehensive all-ops testing.

Aug 11, 2026

Code maintenance release addressing specific review comments from a previous pull request to ensure code quality and correctness.

Aug 11, 2026

Optimized OpenCL Flash Attention prefill kernels by transposing the K tile in local memory to improve performance.

Aug 11, 2026

Updated CI infrastructure to target ROCm 7.14, adopting the new TheRock build system and migrating Windows ROCm jobs to use multi-arch wheels.

Aug 11, 2026

Implements support for multi-output backend sampling, allowing the sampling process to be performed on the backend (GPU/CPU) to enable token speculation and improve performance.

Aug 10, 2026

Fixes a bug where CPU affinity masks were being ignored on Android, which could lead to incorrect thread scheduling or performance issues on Android devices.

Aug 10, 2026

Fixes a bug in the `ggml_roll` operation on CUDA and Metal backends. These backends were not validating tensor contiguity, leading to silent data corruption when permuted tensors were used.

Aug 10, 2026

Adds Multi-Token Prediction (MTP) support specifically for the Nemotron model family, enhancing generation capabilities for these models.

Aug 10, 2026

Updates the vendored `cpp-httplib` library to version 0.53.0, a dependency update that may include security fixes or improvements.

Aug 10, 2026

Adds support for the new 'Granite-Switch' architecture, a dense all-attention model with LoRA adapters that can be selected per-token via control tokens using a router attention mechanism.

Aug 10, 2026

Critical bug fix for the model saver that was incorrectly saving MoE (Mixture of Experts) model parameters, causing models like Qwen2MoE to fail loading after a save/load cycle.

Aug 10, 2026

Refactored WebGPU (WGSL) files to simplify the Flash Attention implementation, improving code maintainability and compilation efficiency.

Aug 10, 2026

A quick follow-up release fixing a dispatch bug in the SpaceMiT backend where Q5_0 quantization was missing, ensuring correct tensor operations on specific hardware.

Aug 9, 2026

A maintenance release focused on CI cleanup, disabling specific build configurations like KleidiAI and openEuler, and providing standard binary builds for various platforms including macOS, Linux, Windows, and Android.

Aug 9, 2026

Fixes the `get_info` API response to correctly report the working directory when tools are running in an isolated environment.

Aug 8, 2026