The AI Appliance deeply integrates hardware and software, built on CXL technology and the self-developed SVBM inference optimization engine to run large-parameter models locally at lower hardware cost. It ships with a multi-agent orchestration engine and mainstream model serving, supports RAG knowledge augmentation, keeps all data on-premises and is fully self-developed across the stack, ready out of the box — especially suited for SMB scenarios.
Compute hardware, LLMs, the agent platform and the knowledge-base engine are integrated in a single device, supporting mainstream processors such as Intel and Hygon and leading domestic and international accelerator cards, covering all scenarios.
All Q&A, uploaded files and user records live only on your own machine — an exceptional AI compute experience at a far more competitive price.
Different scenarios differ in concurrent users, model scale and localization requirements. The six solutions below list their target users, runnable models and performance.
Designed for factory-floor operations, emphasizing round-the-clock stability and shop-floor dedicated agents.
Who it fits:Teams working on the factory floor that need high stability, e.g., tobacco and machinery manufacturing.
Models:Runs Qwen3.8-27B efficiently in parallel plus other small-model combinations.
Performance:At least 20–100 concurrent users, with shop-floor dedicated agents on board.
Pricing:Pricing depends on configuration — contact us for a tailored quote.
Brings full-parameter ultra-large models to universities, corporate labs and research institutes — running locally.
Who it fits:Institutions needing extreme performance with few concurrent users (universities, enterprises, research).
Models:Runs the world's most capable open-source model — DeepSeek R1 671B FP8 — plus Qwen3.8-27B.
Performance:At least 2–4 concurrent users.
Pricing:Pricing depends on configuration — contact us for a tailored quote.
Full-stack domestic hardware, meeting the strict demands of government scenarios for stability, accuracy and long context.
Who it fits:Government departments needing very high stability, professional and accurate answers, longer context, and domestic hardware.
Models:Runs the best value-per-parameter model — Qwen3.8-27B — plus mixed small models.
Performance:50–100 concurrent users, with our proprietary government agent.
Pricing:Pricing depends on configuration — contact us for a tailored quote.
A creative-productivity machine for small and mid-sized teams, supporting text/audio/video omni-modal workflows.
Who it fits:Small teams on a budget who want multi-user work — group creation, SMBs, advertising and design studios.
Models:Runs Qwen3.8-27B efficiently, supporting text/audio/video omni-modal creation.
Performance:20–30 concurrent users; MaxKB and ComfyUI pre-installed.
Pricing:Pricing depends on configuration — contact us for a tailored quote.
A general-purpose compute host built for AIGC creators and data-center O&M management scenarios.
Who it fits:Tailor-made for AIGC creators (text-to-image / text-to-video) and data-center O&M (OPC community scenarios), covering diverse workloads.
Models:Easily handles Qwen3.8-27B and various multimodal models with efficient parallel computing, releasing AI compute to the fullest.
Performance:Strong and stable for at least 20 users; MaxKB and ComfyUI deeply integrated for seamless knowledge-base Q&A and AIGC workflows.
Pricing:Pricing depends on configuration — contact us for a tailored quote.
Three products ship under this line: Lite (KR-01), Base (KR-02) and Enhanced (KR-03), for fine-grained parsing, retrieval, augmentation and generation of enterprise documents in industry, finance and law.
Who it fits:Knowledge augmentation and decision support for industry, finance and law; deeply optimized for unstructured technical documents, with traceable retrieval results ensuring correct work guidance.
Models:Fully compatible with Qwen3.8-27B and mainstream embedding models, with deeply tuned parallel computing; the built-in self-developed agent knowledge engine supports compound tasks such as troubleshooting guidance and compliance checks.
Performance:Stronger on identical hardware — stable for at least 30 users with each concurrent stream above 40 token/s; multimodal and all-scenario support, with overall accuracy above 95%.
Security:All Q&A, uploaded files and user records live only on your own machine.
As AI and large-model training keep growing, resource scarcity has become an urgent problem. The platform pools resources over high-speed IO, with multiple protocols (PCIe, Ethernet, private) driving resource pooling and sharing — greatly improving resource utilization.
The platform provides a hardware foundation supporting diverse protocols and resource combinations, delivering high-bandwidth, low-latency interconnect for data centers and AI clusters.
Integrated rich high-speed protocol interfaces (PCIe / ETH / Aurora) plus management interfaces fully cover customization needs and application development in specialized domains.
One resource pool expands beyond 2T of memory; memory sharing across multiple hosts; single- or multi-tier expansion networking.
Heterogeneous service systems; firewall and data-acceleration processing; resource pooling over PCIe / CXL connections.
New PCIe / CXL distributed storage; memory pool and SSD media backup; low latency, high bandwidth, high performance.
PCIe / CXL networking; 8 GPU interconnect per resource pool; low latency, high bandwidth, multi-tier switching.
| Interface | Qty | Description |
|---|---|---|
| PCIe CEM connector ports | 8 | Attach resource cards |
| ASM cable ports | 2 | HOST-side, connect multiple HOST servers |
| QSFP28 cable interfaces | 4 | HOST-side, connect multiple HOST servers |
In the era of large AI models, compute supply is shifting from "replacing the whole machine" to "elastic expansion". Our product excels in reusing general-purpose servers, AI compute upgrades, multi-instance GPU inference with high concurrency, and long-context LLM applications.
Attach expansion over the standard PCIe interface, breaking the original server's limits on PCIe lanes, slots, power and cooling — add GPU compute to existing x86 or domestic server platforms without replacing the machine.
External expansion with no machine replacement — beyond the original limits on PCIe lanes, slots, power and cooling, protecting existing IT investment.
Multi-GPU VRAM cooperation with tensor parallelism; parameter sync and KV Cache exchange complete inside the box, lowering TTFT and lifting throughput.
Expanded GPU VRAM holds more KV Cache, supporting longer context windows and higher concurrency.
Standard PCIe interface fits mainstream x86 and domestic server platforms, compatible with leading domestic and international GPU accelerator cards — plug and play.
Specializing in high-speed interconnect chips and system solution design.
Hardware-software integrated, delivery-ready
A hardware platform for multi-protocol, multi-resource combinations
Rich high-speed protocol and management interfaces
