iFlytek released Spark X2.5: A domestic MoE large model that enhances code and agent capabilities
2026-09-07 11:39Models🔥 52.3 heat score
3sources
1days unfolding
52.3heat score
3mentions
SummaryAI generated
On September 7, iFlytek officially released the Spark X2.5 large model. This model adopts the MoE architecture with a parameter size of 293B-A30B, focusing on improving code generation and agent execution capabilities, while continuously enhancing mathematical reasoning and general understanding abilities. The model is trained and推理 entirely using domestically produced computing power, aiming to optimize the challenges associated with training and inference for ultra-long sequences. Currently, Spark X2.5 has been launched on the iFlytek Open Platform for users; its end-side versions, Spark X2.5-4B and 1.7B, have also been made open-source. They support context lengths of up to 1 million Tokens and can be locally deployed for long document processing, code assistance, and data analysis.
iFlytek has released the Spark X2.5 large model, which uses the MoE architecture with a parameter size of 293B-A30B. This model focuses on enhancing the capabilities of code and agents. The model is trained and推理 entirely using domestic computing power, continuously optimizing solutions to the challenges associated with training and reasoning for ultra-long sequences. It has now been launched on the iFlytek Open Platform. Previously, the Spark X2.5-4B and 1.7B end-side models were also made open source, supporting contexts of up to 1 million Tokens and can be locally deployed for scenarios such as long document processing, code assistance, and data analysis.
On September 7, iFlytek officially released the Spark X2.5 large model. This model adopts the MoE architecture with a parameter size of 293B-A30B, focusing on enhancing the capabilities of code and agents, while continuously improving general abilities such as mathematics and understanding. The model is trained and推理 entirely using domestically produced computing power, continuously optimizing solutions to the challenges associated with training and reasoning for ultra-long sequences. It has now been launched on the iFlytek Open Platform. Previously, two end-side models, Spark X2.5-4B and 1.7B, were made open source, supporting context lengths of up to 1 million Tokens and enabling local deployment in scenarios such as long document processing, code assistance, and data analysis.
On September 7, iFlytek officially released the Spark X2.5 large model. This model uses the MoE architecture with a parameter size of 293B-A30B, focusing on enhancing the capabilities of code and agents, while continuously improving general skills such as mathematics and understanding. The model is trained and推理 entirely using domestically produced computing power and has now been launched on the iFlytek Open Platform. Two end-side models, Spark X2.5-4B and 1.7B, have been made open-source, supporting a context length of up to 1 million Tokens and can be locally deployed for scenarios such as long document processing, code assistance, and data analysis.