llama.cpp Releases

Focuses on Qwen3Next model support by registering a new tensor for attention normalization and fixing input assignment issues in the layer processing loop to ensure correct DFLASH functionality.

Jun 30, 2026

This release adds support for bracketed IPv6 literals in URL parsing for the server and common components, improving compliance with RFC 3986 standards. It also fixes URL parsing logic to correctly handle proxy headers and client rebuilds.

Jun 30, 2026

This release addresses a limitation in the CUDA backend where `get_rows_back` failed for tables exceeding 65,535 rows by implementing grid-y clamping and stride adjustments. Standard binaries are provided for all major platforms.

Jun 30, 2026

This release fixes a critical bug in the FlashAttention implementation for the Gemma E4B (4 Billion parameter) model with MTP (Multi-Token Prediction) on the CUDA backend. Standard binaries are included for all supported platforms.

Jun 30, 2026

This release optimizes the Vulkan backend for Asahi Linux by rolling back the BK loop in matrix multiplication, likely improving performance or correctness on that specific platform. Standard binaries are provided for all major operating systems.

Jun 30, 2026

This release introduces support for NVIDIA FP4 precision in the WebGPU backend, optimizing memory usage and performance for specific hardware. It provides standard binaries for all supported platforms.

Jun 30, 2026

This release reverts a previous optimization regarding synchronizations during split compute to address potential issues. It maintains the same comprehensive platform support and binary distribution as the previous release.

Jun 30, 2026

This release optimizes the API by removing duplicate entries for presets and cached models in the `/v1/models` endpoint. It includes standard binary builds for various platforms including macOS, Linux, Android, Windows, and openEuler.

Jun 29, 2026

This release introduces granular control over file inclusion levels via `output.patterns`, a new cwd-relative output path style, and a `--skill-project-name` option. It also includes robustness improvements for compression and fixes for file matching.

Jun 29, 2026

Adds comprehensive support for DeepSeek V4 models, including conversion, graph optimization, and multi-sequence processing.

Jun 29, 2026

Fixes a tooling issue where Tailwind CSS scanning was disabled in ignored worktrees.

Jun 29, 2026

Performs code cleanup by removing unused regex-partial functionality from the common module.

Jun 29, 2026

Adds a new flag to preserve reasoning content when generating chat outputs using Jinja templates.

Jun 29, 2026

Fixes a UI bug where stop sequences and reasoning steps were incorrectly skipped in single-model mode.

Jun 28, 2026

Implements a dedicated parser for MiniCPM5, refactoring tool-call handling and fixing Jinja API compatibility.

Jun 28, 2026

Enhances the Jinja template engine with a debugging flag to output the generated program for troubleshooting.

Jun 28, 2026

Introduces support for DFlash v2 architecture, including a new sliding window attention mechanism per layer type.

Jun 28, 2026

Adds an offline mode to the `llama download` tool and fixes a critical use-after-free bug in the URL-task callback.

Jun 28, 2026

This release focuses on reducing logging verbosity and performing code cleanup, specifically renaming constants and updating common components.

Jun 28, 2026

Significant performance and feature updates to the OpenCL backend including a major rework of Flash Attention kernels supporting various quantizations and precisions.

Jun 27, 2026