TAL needed production-grade AI infrastructure and large-model tooling that could train, validate, secure, and deploy its self-developed MathGPT model quickly enough for real education products, while keeping resource efficiency and ROI under control.
Reported impact
TAL shortened the path from model development to classroom productization, improved learning-device experiences, and gained an AI foundation that can answer most math questions instantly while controlling compute efficiency and operational risk.
Baidu AI Cloud supported TAL's self-developed MathGPT model with a high-performance training and serving stack built around the Baige AI heterogeneous computing platform and the Qianfan large-model platform. Baige supplied the cloud-native AI compute layer for large-scale deep-learning workloads, including cluster operations, task lifecycle management, training and inference acceleration, fault tolerance, and intelligent fault diagnosis. TAL used that environment to start projects quickly, validate base-model choices, and iterate MathGPT efficiently with its own educational data and business feedback. On the model-development side, TAL emphasized Qianfan's full-stack workflow support, efficient supervised fine-tuning, plugin extensibility, and security controls, while Baidu also contributed a lifecycle-oriented large-model security solution derived from ERNIE practices. The resulting platform let TAL move from training and validation into real product deployment on the xPad learning device, where MathGPT powers interactive math help and other AI education tools.
Educational data and business feedback are used to validate and fine-tune MathGPT on Baige and Qianfan, accelerated training and inference pipelines prepare the model for production, and the tuned model is embedded into xPad learning workflows to answer math questions, generate AI explanations, and support other study tools.
Baidu AI Cloud used the Baige AI heterogeneous computing platform to build high-performance AI infrastructure for TAL's self-developed MathGPT model. (original: 百度智能云联合大模型落地教育场景先行者好未来，运用功能丰富、性能卓越的百度百舸·AI 异构计算平台，为好未来自研“九章大模型（MathGPT）”提供关键支撑。)
Baige provides cluster operations, lifecycle task management, training and inference acceleration, fault tolerance, and intelligent fault diagnosis, with effective training time above 98%. (original: 百度百舸·AI 异构计算平台...为模型算法专家和运维专家提供全面的集群运维支持和任务全生命周期管理，同时具备训练 / 推理加速、故障容错、智能故障诊断等高级功能，有效训练时长可达 98% 以上。)
MathGPT on the xPad can answer 80% of math questions immediately, upload a human explanation video within one hour for unanswered questions, and generate an AI video explanation within 20 minutes. (original: 基于九章大模型（MathGPT）的“数学随时问”，可以实现对 80% 的数学题即问即答，暂时不能回答的最快可在一小时内上传真人讲解视频、20 分钟内生成 AI 视频解析。)
Qianfan is the one-stop large-model development and service platform that TAL focused on for full-stack workflow support, efficient SFT tuning, plugin extensibility, and business security. (original: 好未来在使用过程中重点关注千帆提供全栈全流程能力、高效的 SFT 微调能力、灵活支持更多的插件能力、业务安全性等。)
The project gave TAL the conditions and technical support to validate, iterate, and upgrade the MathGPT foundation model quickly in the middle phase of the program. (original: 在项目中期，利用百度智能云在百度百舸·AI 异构计算平台上的运行实践...快速完成了九章大模型（MathGPT）底座大模型的快速验证、迭代与升级。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Baidu AI Cloud Sample Oct 13, 2025 Kaoshibao (Shanghai Juxian Network Technology Co., Ltd.) Kaoshibao needed to use new AI interaction and content technologies to improve personalized learning, lower content-production costs, and make very large question banks easier to ingest, search, and explain. "Zero-basis" users improved certificate pass rates 35%. Intelligent Search and Generation Component · Multimodal Recognition · Qianfan AppBuilder Baidu AI Cloud Sample Jun 23, 2025 Tongji University Tongji University needed a domestic AI stack with stronger computing and model-tooling support for embodied intelligence, multimodal models, autonomous perception and multi-agent research. BaiGe AI Heterogeneous Compute Platform · Kunlun P800 · Qianfan AppBuilder Baidu AI Cloud Sample May 15, 2025 Lanxum Education (立思辰) Lanxum wanted to make Chinese-language learning more engaging and personalized, but it needed a structured way to connect teaching content, teacher knowledge, dialogue, and student guidance instead of relying on static… Chinese Teaching Research Knowledge Graph · Intelligent Voice Dialogue System · Virtual Teacher 