Spark-X2.5 is an open model series from XHToken (branded SparkLLM) comprising two compact, general-purpose language models, Spark-X2.5-4B and Spark-X2.5-1.7B, aimed at everyday conversation, writing, translation, reasoning, coding, tool use, and agentic workflows across more than 200 languages. The models use a hybrid attention architecture that pairs one full-attention layer with three sliding-window attention layers, which the README says substantially cuts the compute overhead typical of long-context models while still natively supporting a context window of up to 1,048,576 tokens. Both sizes were trained on Huawei Ascend clusters using large-scale reinforcement learning and post-training techniques the team calls MOPD, and are positioned for deep integration with agent harnesses such as Codex, Claude Code, OpenClaw, and Hermes. The README's own benchmark table shows mixed results against Qwen3.5 and Gemma4 models, several of them larger: Spark-X2.5-4B leads notably on agentic evaluations such as tau3-bench (30.4 vs. Qwen3.5-9B's 9.3) and BrowseComp (40.9 vs. 8.3), and on SWE-Bench Multilingual (53.3 vs. 43.3), while trailing on knowledge-heavy benchmarks such as GPQA (67.4 vs. Qwen3.5-9B's 77.2) and SWE-Bench Verified (41.6 vs. Qwen3.5-9B's 53.1). Weights are mirrored across six channels, Hugging Face, ModelScope, Modelers, Ollama, SCNet, and AtomGit, so users can pick a source suited to their region or hardware, and FP8/INT8 quantized checkpoints ship alongside the base weights for both sizes. Inference runs via vLLM and SGLang, including Ascend NPU deployment through a companion Spark-plugin package, llama.cpp, with native Spark2_5ForCausalLM support added upstream, a companion Spark-MLX-LLM runtime for running the original Hugging Face checkpoints on Apple silicon, Linux CPU, or CUDA without GGUF conversion, plus Ollama and LM Studio for local serving with custom tool-call and reasoning parsers. Fine-tuning is recommended via a Spark-adapted fork of LLaMA-Factory. The repository and model weights are released under the Apache License 2.0, and the project also maintains Slack, Discord, and WeChat community channels alongside its GitHub issue tracker.