llama.cpp Releases

This release fixes thread and block count calculations in quantized copy kernel launches for the SYCL backend, significantly improving throughput for certain quantization types.

Aug 17, 2026

This release adds support for AdamW and SGD optimizer steps in the SYCL backend, along with documentation updates.

Aug 17, 2026

This release removes some instances of the ggml_concat operation as part of refactoring efforts.

Aug 16, 2026

This release refactors the handling of content type capabilities (supports_string_content and supports_typed_content) in the chat module, improving detection logic and adding normalization.

Aug 16, 2026

This release adds a safety check to ensure LoRA tensor data is within the bounds of the file to prevent potential crashes from malformed LoRA files.

Aug 16, 2026

Major feature release adding support for the Kimi-K3 text model, which introduces novel architectural features like latent MoE, situ activation, and MLA output gates. It also includes a dedicated chat format for K3 and MXFP4 quantization support.

Aug 15, 2026

This release redesigns the server's yield-to-queue thread model by moving speculative processing into the worker thread and swapping the worker/main thread design for improved concurrency.

Aug 15, 2026

A minor maintenance release that updates the underlying BoringSSL vendor library to version 0.20260813.0.

Aug 15, 2026

This release adds support for loading MTP (Multi-Threaded Processing) assistant models using the --models-dir flag. It includes updates to presets for strict MTP model prefix checking and removes support for the eagle3 model.

Aug 15, 2026

Bug fix release that adds validation for GGUF array types before reading, preventing potential errors with malformed files.

Aug 15, 2026

Enhances Vulkan backend performance for Intel Xe hardware with optimized shared memory operations and reshape functionality.

Aug 15, 2026

Deprecates and migrates old memory mapping flags to a unified --load-mode argument, improving consistency across the codebase.

Aug 15, 2026

Adds support for MiniMax Text and M1 models with specific optimizations including logits masking and recurrent state management.

Aug 15, 2026

Various bug fixes for multi-token and common modules, addressing stability issues without introducing new features.

Aug 15, 2026

Performance optimization release that fixes quadratic cost issues in jinja template string processing, improving efficiency for template operations.

Aug 14, 2026

Adds support for OpenAI Chat Completions reasoning_effort parameter, making it available to jinja templates for better model control.

Aug 14, 2026

A sync release that consolidates ggml changes without introducing new features or breaking changes.

Aug 14, 2026

Implements recurrent state rollback support for SSM (State Space Model) operations in nemotron models, enhancing performance for certain model architectures.

Aug 14, 2026

This release adds support for virtual integrated GPU devices in llama.cpp, expanding hardware compatibility options for various platforms.

Aug 14, 2026

The server now allows accessing `/metrics` and `/slots` endpoints during `llama_decode()` execution, enabling real-time monitoring.

Aug 14, 2026