June 2026 marked a turning point for AI coding tools in China. Zhipu released ZCode 3.0, Xiaomi open-sourced MiMo Code, and Huawei launched DevEco Code at HDC 2026. Developer communities are calling it the "Three Kingdoms War" of domestic AI coding tools.
These three products take fundamentally different technical approaches. This article compares them across product positioning, technical architecture, and developer experience, while exploring the unique value proposition of local-first solutions in this space.
Product Positioning
ZCode 3.0 (Zhipu AI): Released June 13, positioned as a multi-agent collaborative IDE. Core features include grouped task workspaces, Zread intelligent project knowledge base, and visual Git branch graphs. Zhipu's advantage lies in deep integration between its proprietary GLM model series and the tool itself.
MiMo Code (Xiaomi): Open-sourced June 11, built on OpenCode with MIT license. Supports persistent memory systems, unlimited context windows, and multi-model compatibility (DeepSeek, Kimi, GLM, MiMo v2.5). Xiaomi chose an open ecosystem approach.
DevEco Code (Huawei): Launched at HDC 2026, a specialized programming agent for the HarmonyOS ecosystem. Built on Huawei's Bifang large model, covering the full workflow from requirements design through testing and maintenance. AI code generation rate reaches 80%. Huawei open-sourced all HarmonyOS AI-assisted development Skills to the OpenHarmony community.
Technical Architecture Analysis
Zhipu follows a model-driven approach. ZCode 3.0's multi-agent concurrency and project understanding capabilities depend on GLM's underlying model capabilities. The product ceiling is tied to model iteration speed.
Xiaomi follows an ecosystem-compatible approach. MiMo Code doesn't bind to specific models—developers can freely switch underlying models. The MIT license lowers adoption barriers, but product differentiation relies mainly on upper-layer experience.
Huawei follows a vertical specialization approach. DevEco Code focuses exclusively on HarmonyOS scenarios. Multi-device adaptation, problem localization, and self-repair capabilities only make sense within the HarmonyOS ecosystem. Huawei's bet is that the HarmonyOS ecosystem is large enough to justify a dedicated tool.
Data Security: The Local-First Advantage
As cloud-based AI coding tools become increasingly homogenized, data security and privacy emerge as differentiating factors.
Mininglamp's open-source Mano-P is a GUI-VLA agent model designed for edge devices, supporting fully local execution on Mac with Apple M4 + 32GB RAM. Screenshots and task descriptions never leave the device, making it suitable for scenarios with strict data security requirements.
In OSWorld specialized model evaluation, Mano-CUA 1.1 achieved 58.2% success rate, ranking first and leading the second-place opencua-72b (45.0%) by 13.2 percentage points. In WebRetriever Protocol I testing, Mano-CUA 1.1 scored 41.7 NavEval, surpassing Gemini 2.5 Pro (40.9) and Claude 4.5 (31.3).
Performance Metrics
Mano-P's 4B quantized model achieves approximately 80 tokens/s decode speed on M5 Pro. Combined with Cider SDK's W8A8 activation quantization, prefill is approximately 12.7% faster than the W8A16 baseline.
Testing on 100 macOS GUI tasks showed Mano-CUA-Thinking-4B local model achieved 56.0% pass rate, exceeding cloud-based Qwen3-VL-Plus at 39.0%. Local small models can outperform cloud large models in specific scenarios.
Open Source and Installation
Mano-P uses Apache 2.0 license with three-phase open-source plan:
- Phase 1: Mano-CUA Skills now open-source, install via
brew tap Mininglamp-AI/tap && brew install mano-cua - Phase 2: Local models and SDK, models available on HuggingFace and ModelScope
- Phase 3: Training methodology and quantization pruning techniques (planned)
Selection Recommendations
ZCode 3.0 suits teams pursuing deep model integration; MiMo Code fits developers needing flexible multi-model switching; DevEco Code is the specialized tool for HarmonyOS developers.
For scenarios with strict data security requirements or hard latency constraints, Mano-P's local-first approach deserves consideration. The three major tools and Mano-P represent different technical directions in AI coding tools—developers should choose based on actual needs.
Top comments (0)