As large model technology shifts from competing on parameter scale to achieving large-scale commercial deployment, the criteria for evaluating the computing power market are changing significantly. On one hand, the explosive growth in daily Token invocation volumes has made cloud inference costs and computing power economic efficiency core metrics; on the other hand, driven by millisecond-level response latency requirements, data privacy compliance, and the rapid improvement of small-parameter models at the edge, the industry consensus is that computing power is "sinking" to business sites, creating a layered computing architecture.
Against this backdrop, how to lower the barrier for adaptation across heterogeneous hardware platforms and improve actual delivery and deployment efficiency has become an engineering challenge for domestic AI chips moving toward high-value applications.
At the recently held Iluvatar CoreX 2026 Media Open Day, the company publicly demonstrated the Agentic AI layout of its software stack and the live operation of the "Tongyang TY1200" edge computing terminal. Interactions at the event revealed that domestic general-purpose GPU chips and supporting software are gradually moving beyond basic functional testing to address real-world engineering issues related to cloud and edge deployment, adaptation, and optimization for customers.
Image | Iluvatar CoreX 2026 Media Open Day site; Source: Iluvatar CoreX
In terms of cloud computing power, Iluvatar CoreX showcased third-party benchmark data for its next-generation general-purpose GPU Tiangai 300 (equipped with 144GB of memory). Under typical PD (Prefill/Decode) disaggregated deployment, the Tiangai 300 achieved an end-to-end throughput per card equivalent to 1.05 times that of the NVIDIA H200. In long-context scenarios with 64K tokens using the GLM-5.3 model, its MoE throughput reached 1.15 times that of the H200.
However, as heterogeneous division of labor becomes more refined, the engineering workload borne by the software stack increases exponentially. To reduce migration and optimization costs for customers, Iluvatar CoreX introduced Agents and Skills (skill libraries) into its software stack and deeply integrated FlyDSL, aiming to transform traditionally manual optimization tasks into reusable processes.
According to Iluvatar CoreX representatives, regarding environment and service deployment, the software stack organizes installation and model deployment methods into agent tools, reducing the total time for environment preparation and service deployment in large model PD disaggregation scenarios to 10–20 minutes. In performance bottleneck analysis, the software stack assists AI in identifying compute, memory access, or scheduling bottlenecks based on real runtime data, shortening the tuning cycle to under 2 hours. Furthermore, through collaborative code generation and testing involving Agents and FlyDSL, compilation speed can be increased by over90%; work that previously required 10 person-days can now be completed by one person within hours.
The event also demonstrated a cross-vendor AFD (Attention-FFN Disaggregation) heterogeneous separation solution. In this scheme, the Tiangai 300 handles the full Prefill computation and Attention computation during Decode, while partner vendors' XPUs handle FFN expert network computations. The two types of chips exchange data under unified scheduling, successfully reducing Decoding inference latency by 60% in tests.
Additionally, leveraging the trend of "edge deployment where computing power stays close to data," Iluvatar CoreX extended its accumulated capabilities in model compression, quantization tools, and hardware-software co-design from the cloud to the edge, launching the "Tongyang (TY)" edge product matrix, which includes both computing modules and computing terminals.
The flagship edge terminal, Tongyang TY1200, provides 300 TOPS of INT8 dense computing power, equipped with 32GB of system memory and 32GB of HBM2e graphics memory. In benchmarks using large language models such as Qwen3.8-27B and Qwen3.6-35B-A3B, the TY1200's performance reached 1.2 times that of NVIDIA's edge chip Thor T5000. In tests using NVIDIA's open-source world foundation model Cosmos-R2, the measured performance was 1.58 times and 1.80 times that of the Thor T5000, respectively. Meanwhile, the TY1200 is the first domestic edge product to achieve MuJoCo-Warp-based GPU-accelerated physical simulation at the edge. At the scene, a3D-printed bipedal robot duck (MicroDuck) performed offline physical simulation, policy training, and action verification on a local terminal without internet connectivity.
Currently, the Tongyang series has been deployed locally in multiple industries. In text-to-video scenarios, the TY1200 generates a 5-second video at 1056×608 resolution locally in an average of 6 minutes. Security-sensitive tasks such as legal contract review and investment BP analysis can be processed entirely offline using locally deployed large models. In emergency management, cross-validation between edge CV and multimodal large models improved fire and smoke recognition accuracy from 70% in traditional solutions to over 95%. Additionally, the series has been applied in vertical AI agent report generation for energy storage cell research and has been deployed at the Shanghai Sunshine Rehabilitation Center to provide guidance and navigation services for visually impaired individuals.
Toward the end of the event, Iluvatar CoreX previewed the next-generation edge product, Tongyang TY2000, whose nominal dense computing power will reach 600T, further expanding the scope for offline deployment, elastic scaling, and multi-machine interconnection performance in edge products. With efficient intelligent production in the cloud, broadened real-world applications at the edge, and accelerated iteration via software, a tripartite synergy is forming an autonomously evolving computing architecture, providing engineered technical support for the scalable implementation of AI across diverse industries.