New Model Support and Experimental Implementations in llama.cpp

The open-source inference framework llama.cpp is seeing new developments through both official pull requests and independent community experiments. Recent activity includes the integration of new compact language models and unverified attempts to expand Mixture-of-Experts (MoE) capabilities.

Integration of Spark-X2.5 Models

A new pull request submitted by user KnightYao to the ggml-org/llama.cpp repository introduces support for the Spark-X2.5 series. The implementation covers the Spark-X2.5-4B and Spark-X2.5-1.7B models, which are described as compact, general-purpose language models designed for efficiency and accessibility.

According to the technical documentation provided by XHToken, these models utilize a hybrid attention architecture. This design combines one full-attention layer with three sliding-window attention layers, a configuration intended to reduce computational overhead while supporting a native context window of up to 1 million tokens. The models are reported to support over 200 languages and are optimized for tasks including reasoning, coding, and agentic workflows.

The Spark-X2.5 models were trained on Huawei Ascend clusters using large-scale reinforcement learning and post-training techniques such as MOPD. In terms of hardware compatibility, the models are designed to run on various platforms, including NVIDIA, Huawei, and Hygon. They are compatible with several inference frameworks, including vLLM, SGLang, and MLX, and can be deployed via platforms such as Ollama and LM Studio.

Unverified MoE Expansion Experiments

In a separate development reported on Reddit, a developer has announced a custom branch of llama.cpp aimed at supporting “Expert expansion” for Mixture-of-Experts (MoE) models. The developer, vagrillo, claims to have built this branch with the assistance of GLM 5.3 Flash and reports that initial testing on Apple’s Metal framework shows improved performance compared to previous versions.

The project, hosted on GitHub under the “moex-expansion” documentation, seeks feedback from users on other hardware platforms and different model types. However, it should be noted that these claims remain unverified and lack independent technical confirmation. Discussion surrounding the implementation on community forums has been polarized, with some users questioning the technical substance of the documentation provided.