llama.cpp v0.4.1 Released with Breaking Changes
At a Glance
| Item | Value |
|---|---|
| Repository | ggml-org/llama.cpp |
| Version | v0.4.1 |
| Published | 2026-09-15 |
| License | MIT |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
llama.cpp v0.4.1 has been released. llama.cpp is an inference engine implemented in C/C++ for open-weight models that enables local LLM (Large Language Model) inference.
The most important change in this update is the reorganization of arguments related to model loading methods. Specific load options that were previously available have been completely removed and consolidated into management via --load-mode. Users who specify these options directly in existing startup scripts or configuration files will find that models fail to start after the update, so updating the arguments is required.
Breaking Changes & Deprecations
Model-loading-related arguments that had long been deprecated have been completely removed in this release.
| Old Option | New Option |
|---|---|
--mmap / --mlock / --direct-io |
Value for --load-mode |
Users who have been starting models by directly specifying these options need to migrate to the new format.
Additionally, the following changes have been made for API users:
| Function Name | Change Details |
|---|---|
llama_sampler_chain_n() |
Return type changed from int to int32_t |
This change is intended for API consistency and has no impact on ABI or behavior across all currently supported platforms.
Key Changes
Support for New Model Architectures
Support for several new model architectures has been added.
- Maple 20B-A1B: Added support for DeepGrove’s ternary MoE architecture. It uses ternary weights via TQ1_0/TQ2_0, and currently only supports CPU.
- Tencent Hy 4 (
hy_v4): Added architecture support for Tencent’s Hy 4 preview model. This is a DeepSeek-V3 style MLA + MoE model. - Spark2.5: Added support for GGUF conversion, architecture registration, tokenizer, and computation graph for Spark2.5 models. This model uses fused QKV projection and head-wise sigmoid attention output gate.
Model Conversion and Logging Feature Additions
- QKV Tensor Fusing: Added the
--fuse-qkvflag to fuse Q/K/V tensors when converting from Hugging Face to GGUF. This aims to improve efficiency in the conversion process. - Introduction of Structured Logging: The
--log-jsonlflag enables log output in a structured JSONL format, making system monitoring and analysis easier.
Processing Improvements and Bug Fixes
- JSON Processing Revamp: Internal JSON schema representation has been refactored. This improves the handling of structured outputs from models.
- Server Stability Improvements: Fixed an issue in
llama-serverwhere LRU (Least Recently Used) would hang when multiple requests occurred for the same model. Reliability in multi-request environments has been enhanced. - Multimodal Feature Fixes: Fixed a bug related to speculative decoding after image input.
Supported Models and Hardware
This release adds support for the following new model architectures and quantization formats:
- Maple 20B-A1B: Support for DeepGrove’s ternary MoE architecture. Uses ternary weights via TQ1_0/TQ2_0, and CPU inference is currently available.
- Tencent Hy 4 (
hy_v4): Architecture support for Tencent’s Hy 4 preview model. Compatible with DeepSeek-V3 style MLA + MoE models. - Spark2.5: Support for GGUF conversion, architecture, and tokenizer for Spark2.5 models.
- Qwen3-Next / Qwen3.5: Now explicitly writes
recurrent_layersmetadata during conversion, improving conversion accuracy.
On the hardware side, changes have been made such as disabling lazy tensor loading by default on iGPUs.
How to Update
To update llama.cpp, follow the steps below.
If building from source, clone the repository and build:
git pull
make
Additionally, support for PCH (Precompiled Headers) and unity builds has been added to reduce build times.
Releases Since Our Last Article
Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.
Sources
- ggml-org/llama.cpp v0.4.1
- PR #28334: args: remove mmap/mlock/dio flags from arg parser
- PR #27000: llama: add Maple 20B-A1B ternary MoE architecture (CPU)
- PR #28127: Model: add Tencent Hy 4 (hy_v4) preview architecture support
- PR #27868: [Model] Support for Spark2_5ForCausalLM implementation
Update History
- 2026-09-20: Rewrote the article from re-collected sources and restored it from draft to published.