iFLYTEK Launches Spark X2.5; Tests Show Agentic Work From Financial Reports to Website Design

QbitAI said the launch follows iFLYTEK's earlier open-sourcing of two smaller Spark X2.5 models, a 4B and a 1.7B version, both capable of natively handling 1 million-token contexts on local devices. Community developers quantized the models, moved them to Mac, T4 and WebGPU environments, and connected the 4B model to an agent workflow, helping it reach No. 1 on Hugging Face's trend chart. The cloud API, now promoted at half price, natively supports Anthropic and Responses protocols and can be used with OpenClaw, Claude Code, Codex, Hermes, Grok Build and Pi Agent.

In its first test, QbitAI connected the API to Codex and asked the model to turn a Mid-Autumn greeting into a 3D particle webpage with hand gestures. The generated page let users open their hand to scatter particles like osmanthus flowers, clench a fist to gather them into a full moon, and use single- or two-finger gestures to cycle the particles into rotating 3D phrases such as Mid-Autumn blessings and family-reunion wishes. QbitAI said the interaction was smooth and could be adapted for birthdays or anniversaries.

The second set of tests focused on document-heavy work. Given Nvidia's 61-page Q2 earnings report, Spark X2.5 extracted key figures and produced nine charts plus an investment-risk brief, with reported figures for revenue, profit and net income matching accurately. In another task, two studies from Nature and Science were supplied along with one news article and 20 simulated survey responses, and the model was asked to draft the empirical section of a paper on whether heavy AI users are more productive researchers. It produced a qualified conclusion that evidence supported such a direction and added charts, while also warning that the research sample provided was too small and lacked controlled variables. In an e-commerce task, QbitAI intentionally inserted wrong revenue numbers and submitted records as Word screenshots; Spark X2.5 found the hidden statistical errors and converted its analysis into a PDF.

For the final test, QbitAI asked the model to design an official site for the original fantasy game Ashes of the Fallen God. Spark X2.5 generated a gothic underground-city page with navigation, poster, gameplay announcements and compliance modules. The report highlighted a level and boss section that tied maps and enemies to the game's polluted-god lore instead of listing content in isolation. The report said solo developers without design or front-end teams could use the model with image-generation tools to build a launch-ready site.

The report places Spark X2.5 against a broader shift in the industry from conversational AI to agents that finish and deliver work. It also notes that iFLYTEK has run multiple Spark generations on domestic computing infrastructure since deploying its Feixing-1 domestic AI compute platform in 2023. The report says iFLYTEK goes beyond inference adaptation by keeping training, reinforcement learning, iterative updates and inference on domestic platforms, creating a closed loop that returns real-world deployment problems to the training side. At a 2025 annual-results briefing, iFLYTEK said training efficiency for its MoE model on domestic chips reached 93% of an equivalent-scale A800 cluster, up from 30% in an out-of-the-box state. The report also cited a State Council action plan from August 2025 that calls for more than 70% penetration of next-generation intelligent terminals and agents by 2027, and unconfirmed reports that a leading Chinese open-source model company plans to deploy at least 160,000 domestic AI chips.