iFlytek Launches Spark X2.5, a 293B MoE Model Focused on Code and Agents

According to QbitAI, developers can integrate Spark X2.5 into third-party harness frameworks including OpenClaw, Claude Code, Codex, Hermes, Grok Build and Pi Agent. The launch is accompanied by a limited-time 50% discount on API calls. The report said the model's parameter size is 293B-A30B.

Before the larger model, iFlytek open-sourced two on-device Spark X2.5 models with 4B and 1.7B parameters. Both can natively run a 1 million token context on device, according to QbitAI. After release, community users quantized the models, ported them to Mac, T4 and WebGPU, and connected them to agent workflows for continuous tool calls. The 4B model reached the top of Hugging Face's trending list.

QbitAI tested Spark X2.5 through its API by connecting it to Codex. In the first test, the model generated a 3D particle gesture-interaction webpage for Mid-Autumn Festival. When idle, particles scattered like osmanthus flowers; a fist gathered them into a full moon; gesture "1" made the moon break apart and reform as a rotating 3D "中秋快乐"; gesture "2" rearranged the particles into "阖家团圆"; and an open hand scattered them as stars. QbitAI also created a mini-game called "Moon Palace Express," in which players control space rabbits to collect mooncakes.

In a second test, QbitAI asked the model to analyze NVIDIA's 61-page Q2 financial report and produce charts plus a risk brief. According to the report, Spark X2.5 delivered nine charts and an investment research risk brief after a half-hour morning meeting. QbitAI said figures for FY2027 Q1 revenue, profit and net profit matched, and a future commitments table was clear. The model then handled a research task using two papers from Nature and Science, one news report and 20 sets of mock questionnaire data generated by DeepSeek. Asked to draft an empirical section on whether researchers who frequently use AI have higher work efficiency, the model warned about the data source, then concluded that the evidence supported higher output and self-rated efficiency. It also noted problems such as a small questionnaire sample and missing control variables.

In a later e-commerce data test, QbitAI placed data screenshots in a Word document and deliberately included incorrect revenue figures. According to the report, Spark X2.5 found the statistical bugs in the dataset and rendered a text analysis into a visual PDF. QbitAI said the model performed well in this task.

The third test asked Spark X2.5 to design and generate an immersive official website for an original fantasy action game, "Ashes of the Fallen God." The prompt described a world after the fall of gods, split into floating broken lands, with players surviving among broken temples, gothic dungeons and wilderness polluted by divine remains. According to QbitAI, the model kept the setting on track, delivered a gothic dungeon atmosphere, included standard website elements such as navigation and a homepage poster, and added a level map and Boss section tied to the worldbuilding of "uncontrolled divine power and divine remains pollution." It also included gameplay descriptions and compliance components.

QbitAI framed the tests as part of a shift in the AI industry from chat and code generation to task delivery. The report pointed to the rise of OpenClaw, which let AI call tools and act for users, and to discussions by OpenAI and Anthropic about Harness Engineering, or building workflows so agents can work for hours. It reviewed iFlytek's Spark updates: X1 focused on deep reasoning, X2-Flash added code and agent capabilities, X2-VL added multimodal understanding, and X2.5 put code and agents at the front.

The report also highlighted iFlytek's use of domestic compute. According to QbitAI, iFlytek's work goes beyond inference adaptation, where a trained model is run on domestic chips, and covers training, reinforcement learning, continuous iteration and inference deployment on a domestic compute platform. In 2023, iFlytek launched the "Feixing No.1" domestic compute platform and released the full-chain, independently controllable iFlytek Spark large model. At its 2025 annual results briefing, iFlytek disclosed that its MoE model's training efficiency on domestic compute rose from 30% in an out-of-box state to 93% compared with a same-scale A800 cluster, according to the report.

QbitAI also cited a State Council document issued in August 2025 on deepening the "AI+" initiative, which set a target that by 2027 the penetration rate of new-generation intelligent terminals and agents will exceed 70%. The report said there have been reports that a leading domestic open-source model company plans to deploy at least 160,000 domestic AI chips, though the news has not been fully confirmed. It described iFlytek as an early mover in domestic compute, with training and inference experience accumulated across multiple model generations.