Xiaoyi Claw integrated with Pangu 2.0 Pro, HarmonyOS system-level Agent acceleration

Today, blogger @Adak Feng Lang Ju Xu broke the news: Huawei Xiaoyi Claw has been integrated with the open-source Pangu openPangu 2.0 Pro, with a focus on improving HarmonyOS’s system-level task execution capabilities. It has only been five days since Yu Chengdong officially released openPangu 2.0 at HDC 2026 on June 12. Huawei’s pace this time is much faster than expected.

To be honest, the concept of a “system-level AI assistant” has been discussed for almost two years, from Apple Intelligence to Google Gemini on Android, to OPPO Xiaobu and vivo Blue Heart — almost every company is in the race. But only a few have truly convinced people that “this thing can actually get the job done.” Xiaoyi Claw’s base model replacement this time is a key upgrade for Huawei’s bet on the system-level Agent path.

1. First, clarify: What exactly is Pangu 2.0 Pro?

Let’s recap the information announced at HDC 2026. openPangu 2.0 follows the MoE approach, with two versions:

• Pro version: Total parameters 505B, active parameters 18B
• Flash version: Total parameters 92B, active parameters 6B

Both versions have 512K context windows. This number is not especially groundbreaking — Kimi and MiniMax have already pushed context to the million level — but in device-side + system-level Agent scenarios, 512K is actually quite sufficient. The real bottleneck for long context has never been the number itself, but rather inference cost and effective attention.

The more critical factor is parameter selection. The Pro version’s 505B total parameters and 18B active parameters — about 3.5% — is similar to the strategy used by open-source MoE models like DeepSeek-V3 and Kimi K2. Limiting active parameters to 18B means inference on Ascend chips can keep both single-card memory usage and KV Cache in a manageable range. Huawei officially claims “single-card throughput can reach twice that of other mainstream open-source models.” This claim naturally carries some self-promotion, but given that Pangu’s entire training-to-inference pipeline runs on Ascend with deep graph-computation fusion optimization, a 2× improvement is not exaggerated in certain scenarios.

2. Why Xiaoyi Claw needed to change its base model

We have to start with what Claw actually is.

Xiaoyi Claw is not a chatbot; it’s an Agent form that Huawei introduced earlier this year. Since its debut in March, it has been positioned as the “HarmonyOS Lobster,” aimed squarely at Manus, Zhipu AutoClaw, and similar autonomous task-execution Agents. The biggest difference from traditional voice assistants is: it doesn’t just answer your questions — it does the work for you. Processing documents, making PPTs, replying to emails, organizing files, coordinating multi-device collaboration — jobs only count when they are truly completed.

And the Agent form has very different requirements for its base model compared to a Chatbot:

1. Long-term planning: Even a slightly complex task, like “sort all the invoices from this week’s emails into Excel and send them to finance,” involves dozens of tool calls, state tracking, and exception handling steps.
2. Stable tool usage: Function calling must not fail. A single failure can collapse the entire task chain.
3. Fine-grained instruction adherence: If a user says “edit this picture by removing the background but keeping the reflection on the person,” the model must be able to precisely decompose tasks.
4. Continuous context tracking: States spanning different apps, devices, and time windows must be preserved.

According to iFanr’s “AI Artifact Chronicles” testing at the end of May, the old Xiaoyi had problems precisely in the fourth point — “on the Mate XTs, Xiaoyi frequently lost context, and its answers became incoherent.” This is not something patching can fix — it needs a change of base model.

Pangu 2.0 Pro has targeted optimizations for these points. Huawei’s statement was “better adapted to HarmonyOS, Agent tasks are faster, more accurate, and more efficient” — translated, they likely infused massive amounts of HarmonyOS system call corpus and Agent trajectory data during post-training. Such end-to-end coordinated optimization is something third-party models cannot provide.

3. What does “system-level” really mean?

We need to elaborate here, because this is the core differentiator for Xiaoyi Claw compared to other Agents on the market.

Most mobile AI assistants — including Gemini on Android — are essentially “application-level” solutions. They operate other apps via Accessibility services, screen OCR, and simulated clicks. The problem with this approach is: it’s slow, unstable, and easily blocked by anti-cheat mechanisms. iFanr’s review also noted that in HarmonyOS 6.1, Xiaoyi Help could already “hide the execution process in the background” on native Harmony apps — this is the advantage brought by system-level permissions.

With Xiaoyi Claw plus Pangu 2.0 Pro, in theory it can do several things it previously struggled with:

• Cross-device state sync: Reading half a document on your phone and then continuing editing on your tablet, with AI knowing the context.
• Background batch processing: Saying “compile all business trip reimbursement reports for this month” and having it quietly complete this while you watch videos.
• Multi-app chained calls: From grabbing attachments from email, to archiving in cloud storage, to creating reminders in the calendar, to sending notifications in Feishu — all in one flow.
• Personalized memory: Continuously learning user preferences from local data — something Huawei has previously emphasized at HarmonyOS Smart Home launch — Xiaoyi Claw can “remember user lifestyle and scenario habits.”

The core logic here is: the model layer, system layer, and hardware layer are all developed by the same company. This kind of vertical integration has repeatedly proven powerful in PC and smartphone history — Apple won exactly with this strategy. Huawei is now trying to replicate this playbook in the new AI Agent race.

4. Huawei’s position compared to competitors

Looking across the domestic AI Agent segment in smartphones, the main players are:

• OPPO Xiaobu: Base is its own AndesGPT plus integration of multiple large models, stable approach.
• vivo Blue Heart: In-house Blue Heart large model, good on-device inference, but Agent progress average.
• Xiaomi Super XiaoAi: Key focus at this year’s launch, but real usability still catching up.
• Honor YOYO: Early Agent adopter, strong intent understanding, but smaller base model scale.
• Samsung Galaxy AI: Highly dependent on Google Gemini, unavailable domestically.

Huawei’s differentiator is: it is the only player holding the full set — open-source large model + in-house OS + in-house chip. With Pangu open-sourced, the community can contribute improvements in return — something the others cannot easily match in the short term.

However, frankly, Pangu 2.0 in general capability rankings hasn't yet shown hard data to surpass Qwen3, DeepSeek-V3.5, or Kimi K2.5. Huawei clearly chose a strategic shift away from “ranking competition” toward “winning in system scenarios.” This is a wise choice — the general base model competition is already fierce with Alibaba, Moonshot, and DeepSeek pushing hard.

5. Open-source strategy: 7 major components released starting June 30

The open-source move deserves separate mention.

Huawei’s official announcement stated that openPangu 2.0 will begin releasing 7 major components starting June 30, including:

• Pre-training code (newly open-sourced)
• Post-training code (newly open-sourced)
• Training operators (newly open-sourced)
• Model weights
• Inference framework adaptation
• ……

The last three “newly open-sourced” items are notable — releasing complete pre-training and post-training code plus training operators matches early Llama’s openness, more aggressive than most domestic models that only release weights. Operator-level openness is especially important, allowing external developers to understand real optimization logic on Ascend chips — a genuine investment in Ascend ecosystem building.

This move aligns with Academician Wang Huaimin’s “Crowd Intelligence Paradigm” at HDC 2026 — aiming both to “stimulate crowd intelligence” and to “aggregate crowd intelligence.” Opening training code is stimulation; applying the model in Xiaoyi Claw and HarmonyOS Smart Home is aggregation.

6. What it means for developers

Practically speaking, if you’re a HarmonyOS ecosystem developer, this update offers several points of interest:

1. Agent capability release: Xiaoyi’s open platform previously offered LLM mode, workflow mode, and A2A mode. With Pangu 2.0 Pro integration, A2A (Agent to Agent) mode will see much stronger base capabilities, improving stability when interfacing with external Agents.
2. Freedom for long-context scenarios: 512K context means you can input an entire app’s state and several days of user history for decision-making. Previously limited to 32K/128K, many designs had to be fragmented.
3. Lower Ascend deployment cost: If you already run inference services on Ascend, a 2× single-card throughput improvement directly translates to cost savings.
4. On-device deployment opportunities: The 92B Flash version (6B active) could potentially fit into high-end HarmonyOS PCs for local inference, similar to Apple’s M-series chips running LLMs.

If you’re outside the HarmonyOS ecosystem but want to try the Pangu 2.0 series and compare horizontally with other mainstream models, OpenAI Hub is already tracking Pangu’s open-source model adaptation. You can use a single Key to call GPT, Claude, Gemini, DeepSeek, and Pangu in the same interface — a very convenient tool for model selection evaluation.

7. Remaining questions

We can’t be completely optimistic — several issues still require observation:

1. Which version can run on-device? The Pro version’s 505B total parameters obviously require cloud; Xiaoyi Claw on phones likely uses a cloud API + device-side small model hybrid solution. Huawei has not made the Flash version’s on-device path clear.

2. What is the actual Agent task success rate? This is the hardest metric for Agent products. Manus was criticized for dazzling demos but unstable real success rates. Xiaoyi Claw with a new model needs time and large-scale real data to validate.

3. Privacy and security. System-level Agent permissions are powerful — when something goes wrong, it’s serious. Xiaoyi Claw has already obtained the first authoritative security certification for terminal manufacturers from the China Academy of Information and Communications Technology, but real-world testing is ahead.

4. Ecosystem expansion speed. HarmonyOS native app count is growing, but still much less than Android/iOS. The more apps the Agent can operate, the more valuable it is.

8. In conclusion

Xiaoyi Claw integrating Pangu 2.0 Pro is not a huge news item on its own — it’s just a model base change. But within Huawei’s overall AI strategy, the signal is clear: in-house chip (Ascend) + in-house OS (HarmonyOS) + in-house open-source large model (Pangu) + device-side Agent (Xiaoyi Claw) + smart home (HarmonyOS Smart Home) — this full-stack chain is basically complete.

In the first half of 2026, China’s AI industry shifted clearly from “competing on model parameters” to “competing on scenario implementation.” There aren’t many players who can infuse models directly into the OS and achieve Agent-level control. Whether Huawei’s move is fast and steady will become evident in the second half of the year.

If HarmonyOS’s installed base continues to grow, and Pangu’s open-source strategy continues, this combination might truly deliver something different in the system-level AI race.

---

References

• Huawei Xiaoyi Claw integrates open-source Pangu 2.0 Pro model, improving HarmonyOS system-level task execution capability - IT Home: Original report on Xiaoyi Claw’s integration with Pangu 2.0 Pro
• New generation Huawei HarmonyOS Smart Home released: Upgraded “1+3+N” solution, AI butler Xiaoyi Claw can even raise cats - IT Home: Xiaoyi Claw’s application in HarmonyOS Smart Home scenarios