Xuanwu CLI: OpenClaw on Chinese Chips in 5 Minutes

Clawdbot Global Hype

The AI agent revolution has gone global — and Clawdbot (now widely referred to as OpenClaw) is at the epicenter. With over 130,000 GitHub stars, this 24/7 autonomous “AI Jarvis” has ignited developer enthusiasm worldwide.

But there’s a catch: running it cost-effectively requires local inference — and most mainstream tooling like Ollama only supports NVIDIA CUDA or Apple Silicon. For users of domestic AI accelerators — Huawei Ascend, Moore Threads MUSA, MXChip沐曦, Hygon Suiyuan, Cambricon, Kunlunxin — the path has been rocky: fragmented architectures, steep configuration hurdles, scarce documentation, and no unified abstraction layer.

💡 “Not that we don’t love domestic chips — we just can’t get them to run models.”

That ends today.

🚀 Introducing Xuanwu CLI: The Native Ollama for China’s AI Chips

Launched on February 2, 2026, Xuanwu CLI is an open-source, Docker-native command-line interface designed exclusively for Chinese AI hardware. Think of it as Ollama reimagined for Ascend, MUSA, MXChip, and more — with zero learning curve for Ollama users.

✅ 5-minute setup — xw serve, xw pull qwen3-32b, xw run — done.

✅ Consistent CLI syntax: Fully compatible command set:

xw serve    # Start local inference server
xw pull     # Download quantized models
xw run      # Execute chat/inference
xw list     # List installed models
xw ps       # Monitor active processes

✅ Sub-30s model launch for ≤32B parameter models — thanks to optimized Docker containers and auto-tuned inference engines.

Xuanwu CLI Terminal Demo

🔗 Seamless Integration with OpenClaw & AI Agent Ecosystems

Xuanwu CLI isn’t just a model runner — it’s a drop-in OpenAI-compatible backend. It exposes standard /v1/chat/completions, /v1/models, and /v1/embeddings endpoints.

That means:
– LangChain, LlamaIndex, and VS Code Copilot plugins work out of the box — just change your OPENAI_BASE_URL.
– No code refactoring. No logic changes. Just one config swap.

API Compatibility Diagram

And yes — OpenClaw now runs natively on Huawei Ascend and other domestic chips, powered by Xuanwu CLI as its inference engine.

🤖 Your Ascend server is now your 24/7 AI employee — no cloud tokens, no latency, no billing surprises.

⚙️ Engine-Agnostic Architecture & Auto-Detection

Xuanwu CLI abstracts away hardware complexity via its MLGuider inference engine — a proprietary, performance-optimized runtime supporting:
– Huawei Ascend MindIE (native)
– vLLM (community standard)
– Custom quantization backends for low-bit inference

But the real breakthrough? Automatic chip detection and engine matching. Plug in your Ascend 910B, MXChip G500, or Suiyuan YUANPU — Xuanwu CLI identifies it, selects the optimal engine, configures memory mapping, and starts serving — all without user intervention.

Auto-Detection Workflow

No more:
– Reading 500-page CANN docs
– Compiling from source
– Debugging driver version mismatches

Just ./xw-linux-amd64 serve → done.

🏢 Xuanwu Cluster Edition: Unified Heterogeneous AI Infrastructure

For enterprises and AI data centers, Xuanwu Cluster Edition delivers production-grade orchestration across 10+ domestic chip families, including:
– Huawei Ascend
– Cambricon MLU
– Kunlunxin XPU
– Moore Threads GPU
– MXChip G-Series
– Suiyuan YUANPU

It unifies scheduling, monitoring, API management, and fine-grained metering — turning siloed hardware into a single, scalable, billable AI compute fabric.

Cluster Dashboard

Key capabilities:
– Heterogeneous resource pooling — abstract away vendor lock-in
– Multi-tenant API gateway with rate limiting & quotas
– Real-time usage analytics & cost attribution
– Zero-trust offline operation — all models & data stay on-prem

Cluster Architecture Overview

🌐 Why This Changes Everything

This isn’t just another CLI tool.

It’s a strategic bridge — between China’s world-class AI hardware and the global open agent ecosystem.

Historically, ecosystem dominance has come not from raw FLOPs, but from developer experience. Xuanwu CLI flips the script:
– ✅ Lowers barrier to entry for domestic hardware
– ✅ Enables rapid prototyping and production deployment
– ✅ Accelerates adoption of AI agents on sovereign infrastructure
– ✅ Turns hardware fragmentation into a strength — via unified abstraction

🔑 “We don’t lack compute power. We lacked the thread to string the pearls together. Xuanwu is that thread.”

📣 Get Involved

Xuanwu CLI is 100% open source, MIT-licensed, and actively maintained by Tsingmao AI.

✨ Contribute today:
– GitHub Repository
– GitCode Mirror

⭐ Star it. 🍴 Fork it. 🐛 Report issues. 📥 Submit PRs.

Join the community:

Xuanwu CLI WeChat Group QR

Article originally published by Xin Zhī Yuán (New Intelligence Era), author Ding Hui.