iFLYTEK Open-Sources Spark X2.5 Edge LLMs with 1M Token Context
iFLYTEK has open-sourced Spark X2.5-4B and X2.5-1.7B, becoming the first edge-side general large language models to natively support a 1-million-token context window. Trained on approximately 20 trillion tokens using domestic Chinese hardware platforms, the model weights are now freely available on Hugging Face and GitHub. Bringing a 1M token context capacity to small edge models allows local AI devices to digest long manuals and multi-turn interaction histories without sending data to cloud servers. This significantly improves data privacy, reduces latency, and unlocks practical applications in smart homes, robotics, and local office automation. The models feature a hybrid attention architecture optimized for autonomous agents, code generation, mathematics, and instruction following, with the 1.7B model reaching 90.3% command execution accuracy and an average response time of 0.85 seconds on the Domux smart home benchmark. They natively support popular local deployment engines including vLLM, SGLang, and llama.cpp across hardware from NVIDIA, Huawei, Hygon, and Moore Threads.
## BACKGROUND
Edge LLMs are compact AI models engineered to run on local, resource-constrained hardware like home gateways, personal computers, or mobile devices rather than rely on central cloud infrastructure. A model's context window defines how much text it can maintain in memory during a session, and expanding it usually demands prohibitive memory usage. Hybrid attention architectures combine standard full attention layers with local or recurrent mechanisms to handle massive context windows efficiently on smaller computing chips.