llama.cpp v0.4.1 Released with New Model Architectures and Core GGML Update
llama.cpp has released version 0.4.1, introducing support for new model architectures including Maple 20B-A1B ternary MoE, Tencent Hy 4, and Spark2.5. The release also updates the underlying ggml engine to v0.24.0, adds Kimi-K3 recurrent-state rollback, and introduces a `--fuse-qkv` GGUF conversion flag. This update expands local LLM inference capabilities to emerging ternary Mixture-of-Experts (MoE) and hybrid model architectures while boosting speculative decoding efficiency. Additionally, consolidating memory flags and updating core internal libraries improves performance stability across edge devices and integrated GPUs. Key operational changes include deprecating `--mmap`, `--mlock`, and `--direct-io` in favor of `--load-mode`, as well as disabling lazy tensor loading by default on integrated GPUs. The API updates `llama_sampler_chain_n()` to return `int32_t`, refactors JSON schema handling via `common_schema`, and adds structured JSONL logging capabilities.
## BACKGROUND
llama.cpp is an open-source C/C++ framework designed for high-performance, lightweight LLM inference across diverse hardware platforms using quantized GGUF models. It relies on the ggml tensor library as its core execution engine for low-level matrix computations. Modern LLM architectures frequently leverage Mixture-of-Experts (MoE) routing and recurrent states to achieve lower memory bandwidth usage and faster generation speeds.