Overview
On Nov 15 Huawei Cloud put openPangu-2.1-Pro on its MaaS platform, billed per token. The announcement reads as three sentences: better agents, better coding, higher inference throughput. Two of those point at Huawei's own ecosystem and one points at the bill.
The ecosystem sentences are specific. On coding, the focus is app development in HarmonyOS ArkTS and the generation of Ascend operators. Both share a property: a general model can do part of the work but does not know the house rules. ArkTS is written differently from common front-end frameworks, and an Ascend operator has to line up with a particular chip's memory layout and parallelism. Huawei says it tuned for both, improving the accuracy and usability of generated ArkTS code and lowering the bar for operator work. On agents, the published figure is more than 100,000 specialised examples used to improve skill triggering, tool selection and tool calling, with deep research and HarmonyOS system control named as the target scenarios, meaning the model decides which tool to use and how to finish a chain of operations.
The sentence about the bill is in the price table. Input at 0.8 yuan per million tokens, output at 3.2 yuan, and cached input at 0.016 yuan, with no peak and off-peak split, billed only when generation succeeds. The 0.016 figure carries the most information, one fiftieth of the input price and aimed squarely at workloads that resend the same prefix. An agent running dozens of turns resends its system prompt and tool definitions every turn, so the cache price decides whether long runs make sense economically. If you want to see where this is cheap, skip the 0.8 yuan line and read this one.
On specs: 512K context, with up to 500K input, 384K output and 256K of reasoning chain, plus a switchable deep-thinking mode, function calling, prefix continuation and prefix caching. Access requires a real-name verified Huawei Cloud account, MaaS authorisation, and either a preset service or an endpoint. The quickstart documentation currently lists one region, Southwest Guiyang, and does not say whether others follow the same document.
Treat the word open with care. Huawei calls it an open-source Pangu model, but as of Nov 15 there is no 2.1-Pro repository under the openPangu organisation on Hugging Face, and the parameter count, licence and weight location are unconfirmed. The previous generation, 2.0-Pro, was a mixture-of-experts model with about 505B total parameters and 18B active, under a custom OPENPANGU licence rather than Apache or MIT. The first task in an enterprise evaluation is confirming the licence and whether weights are obtainable, not running a benchmark.
One date needs planning around. The older openPangu-2.0-Flash and 2.0-Pro stopped selling to new users on Nov 6 and go offline on Nov 30. Teams still running them have about two weeks to migrate, and the new model is live in one region, so a cross-region deployment has to be rearranged.
The announcement carries no benchmark scores, only claims of significant improvement. The pieces about wiring into AgentArts, CodeArts and OfficeAce appear in press coverage; Huawei Cloud's own documentation does not repeat them, so do not assume those integrations are live. To find out how much work it can do, bring your own HarmonyOS pages or one or two Ascend operators. That will tell you more than the announcement.
Key Features
- HarmonyOS ArkTS code generation: Tuned for ArkTS and related development tasks to improve the accuracy and usability of generated code, with native HarmonyOS app development named as the target.
- Ascend operator generation: Optimised around the Ascend operator workflow so the model writes operator code for Ascend chips directly, the scarcest skill in Huawei's own stack.
- Skill triggering and tool calling for agents: More than 100,000 specialised examples used to improve when a skill fires, which tool is picked and how calls chain, with deep research and HarmonyOS system control as named scenarios.
- 512K context with switchable deep thinking: Up to 500K input, 384K output and 256K of reasoning chain, with function calling, prefix continuation and prefix caching, and reasoning that can be turned off.
- Cached input at 0.016 yuan: Cached input costs 0.016 yuan per million tokens, about one fiftieth of the input price, with public prefixes detected automatically, which suits long-running agents resending the same system prompt.
- Several model families on one platform: The same MaaS also serves DeepSeek, Qwen, Kimi and GLM, so this is not a single-vendor entry point.
Use Cases
- Building native HarmonyOS apps where the model writes pages and logic to ArkTS conventions
- Generating and tuning operator code for Ascend chips
- System-control agents on HarmonyOS devices that have to choose which tool to call
- Long-running agents that pin the system prompt and tool definitions and lean on prefix caching to hold the bill down
- Evaluating model options for a domestic hardware and software stack by opening one endpoint in the Guiyang region first
Pros
- Coding work points at HarmonyOS ArkTS and Ascend operators, the two places in this ecosystem where people are hardest to find
- Cached input at 0.016 yuan per million tokens pushes the cost of repeated prefixes on long-context agents very low
- 512K context with switchable deep thinking means a simple request does not pay for reasoning it did not need
- Billed only on successful generation, with failures free, which keeps experimentation cheap
- Hooked into Huawei Cloud's existing rate limiting, usage accounting and performance monitoring, so metering does not have to be rebuilt
Pricing
Per token, with no peak and off-peak split: 0.8 yuan per million input tokens, 3.2 yuan per million output tokens, and 0.016 yuan per million cached input tokens. Billing applies only when generation succeeds; failures are not charged and unused capacity costs nothing, with the console invoice as the final word. Access requires a real-name verified Huawei Cloud account, MaaS authorisation, and a preset service or endpoint. The quickstart documentation currently names the Southwest Guiyang region only.
Summary
openPangu-2.1-Pro is a launch with clear trade-offs. Capability points at HarmonyOS ArkTS and Ascend operators, two jobs specific to Huawei's own ecosystem, and the price pushes cached input to 0.016 yuan per million, plainly aimed at long-running agents. Three things need attention: Huawei calls it open source but no weights had landed on Hugging Face as of Nov 15, with parameters and licence unconfirmed; the quickstart covers one region; and the 2.0 series retires on Nov 30, leaving a narrow migration window.
Version History
- Launched on Huawei Cloud MaaS (2026-11-15): Huawei Cloud launched openPangu-2.1-Pro on MaaS, upgrading three areas: agent development, AI coding and inference optimisation. Agents were tuned with more than 100,000 specialised examples for skill triggering and tool calling, with deep research and HarmonyOS system control named; coding targets HarmonyOS ArkTS and Ascend operator generation; inference work covered key operators and the serving path. Specs are 512K context, up to 500K input, 384K output and 256K reasoning chain, with a deep-thinking switch, function calling and prefix caching. Pricing is 0.8 yuan input, 3.2 yuan output and 0.016 yuan cached input per million tokens, with no peak split and charges only on success; the quickstart names the Southwest Guiyang region only. The older 2.0-Flash and 2.0-Pro stopped selling to new users on Nov 6 and go offline on Nov 30.