v2.11.2 v0.4.1

Name: llama.cpp
Author: ggml-org

high risk

ggml-org/llama.cppSep 9, 2026 → Sep 14, 20265 days between releases

What's New

Adds Maple 20B-A1B support and updates ggml to v0.24.0, but introduces breaking API and CLI changes.

Breaking Changes

Changed `llama_sampler_chain_n()` return type from `int` to `int32_t`
Removed deprecated `--mmap`, `--mlock`, and `--direct-io` arguments in favor of `--load-mode`

New Features

Added support for Maple 20B-A1B, Tencent Hy 4, and Spark2.5 models
Added `--fuse-qkv` conversion flag for HF-to-GGUF conversion
Added `--log-jsonl` and `LOG_JSON` for structured JSONL logging
Added `server_subproc` and `waiter` structures for router child-process monitoring

Improvements

Updated ggml library to version 0.24.0 with expanded backend coverage
Improved JSON schema handling and chat parsing
Refactored server child-process handling

Bug Fixes

Fixed MTP context KV cache allocation for DeepSeek2, GLM-MoE, and related models
Fixed GDN normalization from `max` to `rsqrt` for affected Qwen/Kimi/GLM models
Fixed Granite parameter counts and Granite3 MoE unknown parameter counts

Upgrade Notes

Update code calling `llama_sampler_chain_n()` to handle `int32_t` instead of `int`. Replace deprecated CLI flags (`--mmap`, `--mlock`, `--direct-io`) with `--load-mode`.

The Two Releases

v2.11.2

SkillSpector v2.11.2

Sep 9, 2026· 1 month ago

A patch release fixing reference accounting and parser limitations.

Full release notes

v0.4.1

Sep 14, 2026· 1 month ago

llama.cpp v0.4.1 adds support for new models like Maple 20B-A1B and Tencent Hy 4, significantly improves JSON schema handling and logging, and updates the ggml library to version 0.24.0.

Full release notes