Platform Architecture
Xinqi Computing Platform
Developer Access
Developer Access
One Endpoint · One Key
OpenAI-compatible API
Anthropic-compatible API
Works with Standard SDKs
Coding Tools & Agents
Developer Console
MaaS & Routing
MaaS & Routing
KV Hit Rate 90%
Model Catalog
Per-Token Billing
KV-Aware Routing
Session Affinity
TTFT Prediction
Canary & Rate Limits
Inference Acceleration
Inference Acceleration
TTFT ↓20%
Engines
vLLM
In-house Engine
External Engines
Acceleration
PD Disaggregation
Distributed KV Cache
Speculative Decoding
Heterogeneous Orchestration
Heterogeneous Orchestration
Mainstream Chips Supported
Scheduling
Hardware Auto-detection
NUMA-aware Placement
Autoscaling
Self-healing
Rollout & Rollback
NVIDIA GPU
RTX 5090 / 4090
H100 / H200
B300
Domestic GPU
Sugon
Moore Threads
…
NPU
Ascend
…
Computing Center
Computing Center
10,000P · Domestic + NVIDIA
Owned & Partner DCs
GPU Clusters
RDMA Network
Distributed Storage
Liquid Cooling
Observability
Observability
99% Availability
Monitoring & Alerts
End-to-End Tracing
Self-healing
SLA Guarantee
Security
Security
Tenant Isolation
Access Audit
Data Stays On-prem
MLPS Compliance
Core Technology Stack
Heterogeneous Scheduler
Unified management and orchestration of GPUs and NPUs
Disaggregated Inference
Tiered KV cache storage optimizes throughput and cost
Elastic Orchestration
Load-based autoscaling with multi-tenant isolation and quotas
Cluster Operations
Full observability, self-healing and canary upgrades
Achievements
Xinqi Computing Platform White Paper
Covers heterogeneous scheduling, disaggregation and SLO control
Whitepaper coming soonPatents & Software Copyrights
Ongoing work in scheduling, liquid cooling and inference (TBD)
Patent list in preparationR&D Philosophy
Built for large-scale inference, we keep optimizing cost and stability so every token is measurable and reliable.