iFlytek's TokeSpark today announced the release and open-source availability of two general-purpose edge models, Spark X2.5-4B and Spark X2.5-1.7B.
According to the announcement, both models natively support a context window of up to 1 million Tokens, making them currently the only edge models supporting a context window of this scale.
Both models use a hybrid attention architecture and are optimized around core capabilities including agents, coding, mathematics, and instruction following. They were trained end to end on a fully domestic computing platform, using approximately 20 trillion Tokens of diverse data for pretraining.
In a product after-sales scenario, for example, users can import a complete after-sales manual into the model. For complex questions such as “the device failed 10 days after purchase and third-party consumables were used,” the model can correlate provisions across chapters to provide a comprehensive judgment, while maintaining the preceding context and continuing to offer advice when new conditions are added.
For personal office work, Spark X2.5-4B can complete the entire workflow from raw data analysis to delivery of a bilingual report. On the Domux smart home test set, Spark X2.5-1.7B achieved an end-to-end execution accuracy of 90.3% for control instructions, with an average response time of 0.85 seconds. The models are also suitable for robot operation control, target tracking, navigation decisions, and other scenarios, and can be deployed on robot platforms or edge devices.
Spark X2.5-4B and 1.7B support hardware platforms from NVIDIA, Huawei, Hygon, and Moore Threads, and are compatible with mainstream inference frameworks such as vLLM, SGLang, and llama.cpp. They can be quickly deployed through tools including Ollama and LM Studio. Starting today, the model weights are available on platforms such as Hugging Face and GitHub. The corresponding APIs are also live on iFlytek's Spark MaaS platform and are free for a limited time.
On September 7, iFlytek will officially release the Spark X2.5-293B general-purpose model, further upgrading core capabilities such as coding and agents. IT Home will bring you a detailed report at that time.
References:
Hugging Face: https://huggingface.co/collections/XHToken/spark-x25
GitHub: https://github.com/XHToken/Spark-X2.5
iFlytek Spark MaaS platform access: https://maas.xfyun.cn/modelSquare