Huawei's openPangu-2.1-Pro Aims at AI Agents and HarmonyOS

Share

Huawei Cloud's MaaS platform now serves the open-source openPangu-2.1-Pro model, built for AI agents, ArkTS code for HarmonyOS and Ascend operators. Read what changed and the pricing, plus a prompt that has your AI agent compare it against your own tasks.

Machine translation. Read the original

Huawei Cloud put the open-source openPangu-2.1-Pro model on its MaaS model service platform on November 15. This version focuses on three things: AI agent development, AI coding and faster inference. It is especially aimed at HarmonyOS developers. It can write HarmonyOS apps, operate the HarmonyOS system and generate operators for Ascend chips. It is pay-as-you-go, with input priced at 0.8 yuan (RMB) per million tokens.

What is new

Agent skills: Huawei says it used more than 100,000 specialized training items. The main improvements are knowing when to use a skill, choosing the right tool, and calling the tool correctly. The goal is for the model to plan better and act more reliably on complex tasks. The biggest gains are in deep research and in operating the HarmonyOS system.

Coding: The focus is on HarmonyOS app development and Ascend operator generation. HarmonyOS apps use the ArkTS language, which most large models have seen very little of, so the code they write often does not run. This version is specifically tuned to make ArkTS code more accurate and usable. An Ascend operator can be understood as the low-level computing code that runs fast on Huawei's AI chips. Until now, engineers had to write it by hand.

Faster inference: Huawei optimized key operators and the inference pipeline. It says throughput in typical deployments has "greatly improved," but it did not give a specific multiple.

Where to use it and what it costs

The model is available directly through an API on the Huawei Cloud MaaS platform. The platform includes rate limiting, usage statistics and performance monitoring. Later, the model will also connect to several other Huawei Cloud products: the AgentArts agent development platform, the CodeArts coding tool and the OfficeAce office product.

Billing item Price
Input 0.8 yuan (RMB) per million tokens
Output 3.2 yuan (RMB) per million tokens
Cached input 0.016 yuan (RMB) per million tokens

A cache hit costs one fiftieth of normal input. AI agents resend long system prompts and tool descriptions in every turn, so this part becomes much cheaper when it hits the cache. To learn how caching works, see the cache handbook.

What it means for you

  • HarmonyOS app developers: There is now another model trained on ArkTS. You can compare it with the model you use now.
  • Teams building AI agents on a tight budget: Low prices and a cache discount suit agent tasks that make many calls.
  • Regular users: You do not need to try it now. Wait until it reaches Huawei's office and coding products.

How to get started

If you already have a Huawei Cloud account, you can have your AI agent run a small comparison:

I want to try openPangu-2.1-Pro on Huawei Cloud MaaS to see if it fits my project. Please help me in this order:

1. First check the official Huawei Cloud MaaS documentation. Tell me the steps to activate this model and where to get an API key. List the steps I must do on the website myself. Do not log in or pay for me
2. After I get the API key, write a minimal call script that reads the key from an environment variable. Do not put the key in a code file
3. Ask me for the 3 things I most often ask AI to do, such as writing a piece of ArkTS code or calling a tool to look up information. Test openPangu and the model I use now on each of those 3 things
4. Compare them in a table: whether the results are correct, whether they can be used as is, and how many tokens and how much money each one used
5. If the total test cost goes over 5 yuan (RMB), stop and ask me first

After each step, tell me the result and wait for me to confirm.

Testing with 3 real tasks of your own tells you more about whether the model fits you than the benchmark scores a vendor publishes.

Things to note

  • The parameter size, open-source license and repository address have not been verified yet. Before commercial or private use, check Huawei's official open-source notes.
  • Huawei did not publish a specific figure for the "greatly improved" throughput. Your own tests are the real measure of speed.
  • Prices are set by the Huawei Cloud MaaS pricing page, and they may change with promotions.
Done with this one? Mark it to track your path.Sign in
Save it for laterSign in
Questions about this piece?Ask AI about this

Comments (0)

No comments yet. Start the conversation.

Sign in to join the discussion.Sign in