Qingcheng.AI
Qingcheng.AI is an AI-infrastructure and intelligent-computing-systems company focused on making large-model training and inference more efficient, usable, and cost-effective across heterogeneous hardware. Its stated goal is to provide ‘more, faster, better and cheaper’ AI system solutions, empower China’s domestic computing ecosystem, and reduce the cost of deploying large models.
At a Glance
- Large-model developers and AI application companies
- Enterprises requiring private AI deployment
- Domestic AI-chip and server manufacturers
- Cloud providers, supercomputing centers, and intelligent-computing centers
- +2 more
AI Tools by Qingcheng.AI
(1)
Chitu
Open Source LLM Inference Engine
Discussions
No discussions yet
Be the first to start a discussion about Qingcheng.AI
Latest News
National Supercomputing Internet and Qingcheng.AI announced deep strategic cooperation on domestic AI-computing and end-to-end token services.
Chitu v0.6.0 released with the self-contained chitu.run executable for multi-node, multi-instance, and prefill/decode-disaggregated jobs.
Qingcheng.AI presented its full-link token solution centered on Chitu, AI Ping, and Bagualu.
Chitu v0.5.1 released with Moore Threads GPU support.
Products & Services
An end-to-end training and intelligent-computing stack covering compiler, operators, parallelism, quantization, scheduling, orchestration, fine-tuning, model quantization, alignment, and service release. Components include ElmNet, PET, Intelligen, SmartMoE, FastMoE, MixQ, Romeo, and QFactory, with support for bare-metal, Kubernetes, and multi-cloud environments.
An open-source, production-grade large-model inference framework jointly released with Tsinghua. It supports heterogeneous CPU/GPU and multi-node deployments, low-latency, high-throughput, and low-memory optimization, software FP8/FP4 conversion, and major domestic and NVIDIA hardware.
A large-model service evaluation and API intelligent-routing platform that aggregates providers through an OpenAI-compatible API, continuously measures latency, throughput, reliability, and price, and routes requests to suitable providers.
Hardware/software integrated solutions based on Chitu for desktop workstations, all-in-one servers, and server clusters, with private deployment, operations, and optimization services.
Market Position
Qingcheng.AI positions itself as a domestic AI-infrastructure software specialist spanning training, inference, and model-service routing, rather than as a model vendor. Chitu competes with general-purpose inference engines such as vLLM and SGLang by emphasizing domestic accelerator optimization, FP4/FP8 support, and heterogeneous scaling; Bagualu competes in training-system and AI-software-stack optimization; AI Ping competes with model-API aggregators and routing/observability services. The company claims Chitu used 50% fewer GPUs and achieved 3.15x the output speed versus vLLM for a DeepSeek-671B test on an A800 cluster.
Leadership
Founders
Tang Xiongchao (汤雄超)
Co-founder and CEO; PhD in computer science from Tsinghua University (2019). Previously worked at Sangfor Technologies and ByteDance; specializes in performance analysis and optimization, parallel and heterogeneous computing, and cluster-resource scheduling.
Tang Shizhi (唐适之)
Co-founder; a former core member of the Tsinghua supercomputing team and part of the Tsinghua Computer Science high-performance-computing research community.
Shi Tianhui (师天麾)
Co-founder and product vice president; won a national Informatics Olympiad gold medal and came from the Tsinghua Computer Science high-performance-computing research group. He left multiple large-company offers to commercialize the group’s systems research.
Executive Team
Tang Xiongchao (汤雄超)
CEO
Tsinghua Computer Science PhD; previously at Sangfor Technologies and ByteDance; works on performance optimization, parallel/heterogeneous computing, and cluster scheduling.
Jin Jiangming (靳江明)
COO
PhD in computer engineering from Nanyang Technological University; background in high-performance computing, AIoT edge computing, and distributed-system architecture.
Board of Directors
Founding Story
In December 2023, young researchers from Tsinghua University’s Computer Science Department High Performance Computing Institute founded the company to turn years of laboratory work in large-scale training, compiler optimization, parallel computing, and domestic supercomputing into deployable products. The team chose a software-first approach: optimize system software for domestic chips so they can run frontier large models efficiently, rather than merely relying on foreign hardware/software stacks.
Business Model
Revenue Model
Enterprise software and services: private deployment, inference and training optimization, integrated hardware/software appliances, technical operations and support, and AI infrastructure solutions. AI Ping is positioned as an API aggregation, evaluation, and routing platform; public sources do not state its pricing mechanism.
Target Markets
- Large-model developers and AI application companies
- Enterprises requiring private AI deployment
- Domestic AI-chip and server manufacturers
- Cloud providers, supercomputing centers, and intelligent-computing centers
- Financial services and energy-sector organizations
- AI developers and teams consuming model APIs
- Enterprise private deployment of DeepSeek, Qwen, GLM, Kimi, and other large or multimodal models
- Large-model pretraining, fine-tuning, quantization, alignment, and inference
- Production AI services on domestic accelerator clusters
- Cost-sensitive inference where lower hardware requirements and higher throughput matter
- Developer and enterprise selection, monitoring, and routing of model APIs
- AI applications in finance, energy, supercomputing centers, and industrial computing environments
- Zhipu
- A major energy state-owned enterprise
- Shengshu Technology
- Brokerage and energy-sector deployment scenarios