Xunfei Spark X2.5 Officially Released: MoE Architecture with 293B Code and Intelligent Agent Capabilities Significantly Improved

iFLYTEK officially released the Spark X2.5 large model. This model adopts a MoE architecture, with a parameter scale of 293B-A30B, focusing on improving code and agent capabilities, and continuously enhancing general capabilities such as mathematics and understanding.

Spark X2.5 was trained and inferred using fully domestic computing power throughout the entire process, and continues to optimize the challenges of domestic ultra-long sequence training and inference. It is now available on the iFLYTEK Open Platform.

The edge-side model has been released as open source first, supporting a context of up to one million Tokens

Previously, the edge-side models of Spark X2.5-4B and 1.7B have already been released as open source, supporting a maximum context of 1 million Tokens, which can be deployed locally for scenarios such as long document processing, code assistance, and data analysis. With the release of this X2.5 flagship model, iFLYTEK has completed the layout of new versions on both cloud and edge product lines.

Xunfei Xinghuo X2.5 Launch: MoE Architecture Provides Significant Enhancement, Code Writing and Intelligent Agent Capabilities Achieve Leapfrog Upgrades

iFLYTEK released the Xinghuo X2.5 large model, which uses a MoE (Mixture of Experts) architecture. The parameter size reaches 293B-A30B. It has been comprehensively upgraded in terms of architecture design, core capabilities, and domestic computing power compatibility, with a focus on enhancing code writing and intelligent agent (Agent) application development capabilities.

iFLYTEK to Launch Spark X2.5 General Large Model

On September 1st, iFLYTEK will open-source two endpoint general large models, Spark X2.5-4B and Spark X2.5-1.7B. They support a native 1M token long context, with enhanced capabilities in intelligent agent interaction, mathematical computation, and general understanding. These models are designed for endpoint scenarios such as automotive, smart hardware, and the Internet of Everything. On September 7th, iFLYTEK will also release a larger-scale model.

China's Large Model Receives Another Major Upgrade, iFLYTEK to Launch a New Main General-Purpose Large Model

iFLYTEK revealed at its 2026 Semi-Annual Performance Briefing that it will officially launch a new main general-purpose large model based on domestically-produced computing power during this year's 1024 Developer Festival; a phased version will be released in late August for early market and developer experience, with an established technical implementation schedule.

Tencent Hy3 Programming Evaluation Released: Parameters Are Only One Fifth of Competitors, Code Ability is on Par with DeepSeek-V4-Pro

SuperCLUE released the programming special evaluation of Tencent Hy3 language model, comparing it with DeepSeek-V4-Pro and others. Hy3 uses MoE architecture, with a total parameter count of 295B and only 21B activated, supporting a 256K context length, and is claimed to be the strongest in the Mix. The results show that despite far fewer parameters, its performance was surprisingly excellent. Evaluated from four dimensions, it balances performance and cost, focusing on real coding scenarios for Chinese programmers.

Tencent Hunyuan Hy3 Launch: Significant Improvement in Agent Capabilities and Product Experience

Tencent Hunyuan Hy3 is launched, with intelligent level comparable to 2-5 times the parameter flagship models, lower pricing, and improved stability and cost-effectiveness. It has been integrated into multiple business scenarios, with API available on Tencent Cloud TokenHub and overseas platforms. The model realizes the integration of fast and slow thinking, bringing a qualitative improvement in the practicality of intelligent agents.