Public developments and original analysis · Announcement date: · Sources
Huawei Cloud MaaS added openPangu-2.1-Pro, version 20261115, to its November 15 release record. It supports a 512K context, function calling and prefix caching. The record lists preset service in Southwest-Guiyang 1. For teams repeatedly supplying project material or long background information, the important condition is separate billing for cached and ordinary input: the cost of a long task depends on what is actually reused.

The China-site table lists CNY 0.8 per million uncached input tokens, CNY 0.016 for cached input and CNY 3.2 for output, without peak/off-peak pricing for this model. All three contribute to the bill. Applying the lowest rate to every input token, or borrowing another model’s nighttime discount, understates cost. Tasks generating long answers can also spend much of their budget on output.
Which tokens receive the cache rate
The billing documentation charges uncached input at its ordinary rate, cached input at the cache rate, then adds output. Prefix recognition is automatic, but the hit rate is uncertain. The cache guide measures the share of input tokens reused. A large number of repeated requests does not establish that most of their text was cached.
Consider hypothetical cumulative usage over a reporting period: one million input tokens and 100,000 output tokens. With no cache hits, the listed rates give CNY 1.12. If 60% of input hits the cache, input costs CNY 0.3296 and output adds CNY 0.32, totaling CNY 0.6496. This arithmetic excludes other services and business-processing costs; it is not a performance measurement. The cumulative usage may span many requests and does not imply that a single request accepts that input volume.
The difference applies only to reused input, not the whole task. Follow-up questions, retries or longer final text can offset input savings with additional usage. Compare all calls needed to finish the same job rather than using the request with the highest cache hit as an average.
Latency needs its own observation. The billing documentation says hits may reduce time to first token, while misses may increase it. Interactive responses and batch document work can justify different arrangements. Cheaper repeated input alone does not establish faster responses for every request.
Organizing stable background material
Prefix caching reuses a repeated beginning, so input construction matters. Stable task instructions and shared reference material can precede the changing object and question. Stability should follow the material’s continued validity; an obsolete rule should not remain merely to preserve a cache hit.
Consider a hypothetical team repeatedly asking different internal document questions about the same product-manual version. The manual and formatting requirements can form shared context, while the question and supplementary evidence vary. Inserting a timestamp or other changing text at the very beginning can prevent that otherwise identical manual from forming an identical prefix.
Fixing more content has costs too. A complete manual still occupies context in each request and irrelevant material can add processing and review work. If retrieval supplies enough relevant passages for occasional questions, carrying the whole manual solely to seek a cache discount may be unnecessary. Compare actual input, hits, output and answer quality before choosing the document scope.
When material changes, correctness comes before reuse. Revised manuals, changed company rules and different user permissions require input to be reconsidered. Internal records also require clarity about cache isolation, retention and deletion. A tariff explains billing; it does not by itself establish how long a record should persist or whose requests may reuse it.
Hosted access and the scope of adoption
The release table marks preset service as supported and public-pool and dedicated-pool deployment as unsupported for this version. That describes the current MaaS offering. Teams requiring dedicated resources or another deployment arrangement should establish a suitable route before evaluating capability, rather than inferring every deployment option from the model name.
Huawei Cloud positions the model for agents, HarmonyOS application development and Ascend operator generation. Those claims help select evaluation tasks, not establish results in an enterprise project. Code should be assessed in its actual toolchain; document work needs checks for facts, references and omissions. A 512K context provides capacity without proving that all long-form material will be used accurately.
The openPangu family name and an available hosted version also need to be distinguished from specific downloadable assets. Self-hosting requires version-matched weights, permissions and runtime arrangements. A cloud tariff establishes none of those. If hosted access meets the need, maintaining an inference environment may add unnecessary work.
Tasks with substantial repeated background, stable document versions and observable usage are candidates for a complete-cost comparison. One-off input, output-heavy work or incompatible deployment requirements provide weaker reasons to switch for a low cache rate alone. Total usage at acceptable quality and the maintenance burden are more useful decision criteria.
Sources
- Huawei Cloud MaaS model release bulletin
Source published: · Accessed:
- MaaS model list and openPangu 2.1 Pro capabilities
Accessed:
- MaaS text-generation billing and cache formula
Accessed:
- Token cache hit-rate guide
Source published: · Accessed:
Mokyun Editorial · neo@mokyun.com
