Iluvatar CoreX says the 2026 China International Industry Fair ran in Shanghai from October 12 to 16. The company brought its Tiangai, Zhikai and Tongyang cloud-edge-device product matrix to the show, laying out a full lineup that runs from flagship cloud GPUs and super node systems to edge inference, computing power modules and computing power terminals, with a focus on large-scale deployments in finance, energy, urban governance, healthcare and industry.

On the eve of the fair, the Communist Party of China Central Committee and the State Council issued the "Opinions on Developing New Quality Productive Forces," calling for full implementation of the "AI Plus" initiative and stronger, more efficient supply of computing power, algorithms and data. Iluvatar CoreX said it will build on its "general-purpose GPU plus in-house software stack" approach to improve the economics of computing power systems and support the development of new quality productive forces.

Cloud: Tiangai 300 Flagship and the IXL144 Super Node

Iluvatar CoreX unveiled Tiangai 300, its new flagship general-purpose GPU, during the World Artificial Intelligence Conference in July 2026. The product was upgraded across chip architecture, key operators, system scalability and software ecosystem, targeting trends in information processing, task execution and open-ended exploration.

Building on the single-chip capabilities of Tiangai 300, Iluvatar CoreX showed the first of its systemized super node designs for the first time: the Tiangai 300 IXL144 packs 144 GPUs and 18 CPUs into a single rack with roughly 10.1TB of memory. By the company's calculations, it cuts the cost per million effective tokens by 40% compared with mainstream international systems.

Edge and Device: The Roles of Zhikai and Tongyang

As AI applications move from the cloud to the field, large-scale training and complex re-verification stay largely in the cloud, while real-time response, data security and on-site closed-loop tasks have to run at the edge or on the device. Iluvatar CoreX built its CIIF showcase around that split, presenting a product matrix spanning cloud, edge and device:

  • Zhikai series: aimed mainly at cloud and edge inference, optimized for latency, throughput and energy efficiency;
  • Tongyang series (launched this year): serves the device side in forms such as computing power modules and computing power terminals, covering physical AI, smart devices and on-site operations.

Extending its cloud general-purpose GPU architecture and software stack to the device side lets devices keep running newer and larger models, the company said, with model accuracy aligned through cloud training, device deployment and cloud-side re-verification. The Tongyang products on display can serve as an "external brain" for embodied intelligence hardware, supplying model execution and real-time inference. In data-sensitive work such as analyzing investment materials or reviewing contract terms, the Tongyang TY1200 computing power terminal can handle processing offline using a locally deployed large model.

Deployment: 30-Plus Industries, 380-Plus Customers, More Than 1,000 Projects

Iluvatar CoreX said the value of computing power ultimately has to be proven by results in production environments. The company now serves more than 380 commercial customers across more than 30 industries, with over 1,000 projects deployed.

Industry: Working with Supcon's TPT time-series large model appliance, the company used Tiangai products to connect the customer's in-house industrial computing engine, running model training, fine-tuning tests and performance validation of proprietary algorithms. It also completed a dual-card inference deployment of a 27B general-purpose large model on vLLM, supplying computing power for process optimization, operating-state prediction and operator assistance.

Finance: Using a domestic GPU computing power platform, the company supported deployment and adaptation of mainstream multimodal large models for UnionPay Cloud, improving response times and concurrent throughput under mixed text-and-image workloads. At a major bank, its software stack supported Monte Carlo simulation algorithms and worked with the CuPy and PyTorch heterogeneous programming frameworks; in a curve interpolation workload covering 1 million rows of data, it ran roughly 4 times faster than a CPU-based setup.

Energy and mining: In geophysical exploration, a unified GPU resource pool now spans the full workflow of data processing, velocity modeling, migration imaging and intelligent interpretation, with Tiangai products deployed at the thousand-card scale. In smart mining, a single server with four Zhikai 100X cards processes 50 channels of 1080P video in real time, combining video decoding, AI recognition and encoded transmission back to base. Paired with Baidu's Yijian platform, the system covers 48 types of work skills across 6 categories, with accuracy of about 85% in flagging rule violations.

Urban governance: Working with Kaiyun Zhilian, the company identifies long-tail events such as road debris and traffic accidents using a small edge model for initial detection and a multimodal large model for re-verification. In testing, the large model reached over 90% accuracy on core event categories, event coverage expanded from 7 categories to more than 16, and the target average detection time is under one minute.

Campus and property management: With Sige Hulian, the company uses existing cameras and edge computing power for rotating patrol recognition, closing the loop from alert to work order. A single device can cycle through 500 video channels in 30 minutes, with alert accuracy up more than 90%. At Xingtong Building, inspection intervals dropped from 4 hours to 30 minutes, a reduction of 83%.

Healthcare: In a partnership with Winning Health, domestic GPUs run the WiNGPT medical large model connected to the WiNEX business system, delivering real-time flags for gaps in medical records, real-time review against 9 categories of health insurance rules, and 24/7 clinical decision support in the ICU.

Software Stack and Developer Ecosystem

Iluvatar CoreX has been building out its general-purpose GPU software stack since 2018, developing close to 100 acceleration libraries and a tool system covering communication, compilation, drivers, quantization and performance analysis. It supports mainstream large model frameworks, models and key acceleration components, with the aim of lowering customers' migration, adaptation and performance tuning costs.

At the fair, developers could also try Iluvatar SkillHub, calling Skills through an Agent to handle model adaptation and accuracy tuning on the Iluvatar computing stack.

What Comes Next

The company said it will update its computing software stack and build an intelligent development suite designed for Agent collaboration, bringing AI into model adaptation, deployment and performance optimization. It plans to keep developing its "general-purpose GPU plus in-house software stack" system, pushing high-quality computing power from single cards to full racks and from the cloud to the device side, for a wider range of industrial use cases.