Spark-X2.5
An open, compact language model series (4B and 1.7B parameters) tuned for coding, agentic tool use, and everyday tasks, with a native context window of up to one million tokens and broad hardware and framework support.
Spark-X2.5 is an open model series with two sizes, 4B and 1.7B parameters, designed to make capable AI practical on modest hardware. The models cover conversation, writing, translation, reasoning, coding, tool use, and agentic workflows in more than 200 languages, using a hybrid attention architecture that combines one full-attention layer with three sliding-window attention layers to natively support context windows of up to one million tokens with reduced compute overhead.
They are integrated with agent harnesses including Codex, Claude Code, OpenClaw, and Hermes, and are compatible with inference frameworks such as vLLM, SGLang, llama.cpp, and MLX, deployable through Ollama or LM Studio, and can be fine-tuned with common training frameworks. They were trained on Huawei Ascend clusters using large-scale reinforcement learning and post-training techniques.
Hardware support spans NVIDIA, Huawei Ascend, Hygon, and other platforms, and the models are also available in a mobile app.