llama.cpp v0.4.1: support for Maple 20B-A1B, Tencent Hy 4 and Spark2.5

The llama.cpp v0.4.1 release came out on September 14, 2026. The update adds support for three new architectures: Maple 20B-A1B — a ternary MoE for CPU, Tencent Hy 4 in preview status, and Spark2.5. In addition, ggml has been updated to v0.24.0, and JSON schema handling, chat parsing, logging, and server child process management have been improved.

What it means

For those running models locally, the main thing is the new architectures. Maple 20B-A1B is described as a ternary MoE model with CPU support, which means it can run without a discrete GPU. Tencent Hy 4 has been added in preview mode: the developer warns that support may be incomplete and will change in future releases. Spark2.5 is another new architecture, details of which are not disclosed in the release notes. The API changes affect developers: llama_sampler_chain_n() now returns int32_t instead of int, and server-common.h adds server_subproc and waiter for monitoring the router's child processes. For model conversion, a --fuse-qkv flag has appeared, allowing Q/K/V tensors to be merged when converting HF to GGUF. The update also fixes bugs in working with Kimi-K3, DeepSeek2, GLM-MoE, Qwen, and Granite. The deprecated arguments --mmap, --mlock, and --direct-io have been removed — --load-mode is used instead. Structured JSONL logging via --log-jsonl has been added. For those who use llama.cpp for inference on their own machine, the update is worth considering primarily for the fixes in existing models and the new Maple architecture for CPU. More on neural network news and tools for local running — in the CPU3D news feed.