Huawei Cloud announced on November 15 that openPangu-2.1-Pro, the latest model in its open-source Pangu family, is now available as a preset API service on its MaaS (Model as a Service) platform. The model first appeared about ten days earlier inside Huawei's Xiaoyi Work agent; the MaaS launch is what opens it up to outside developers. The release notes list version 20261115 with a 512K-token context window — maximum input 500K, maximum output 384K, with chain-of-thought capped at 256K — alongside function calling, prefix completion, prefix caching and a toggleable deep-thinking mode. The service is initially available only in the CN Southwest-Guiyang1 region.
The pricing is the eye-catching part. Huawei Cloud charges 0.8 yuan per million input tokens, 3.2 yuan per million output tokens, and just 0.016 yuan per million cache-hit input tokens, with no peak/off-peak tiers. Its predecessor openPangu-2.0-Pro charged 14.5 yuan per million output tokens for requests under 32K tokens and 17.6 yuan above that, which makes the new output price roughly 18 to 22 percent of the old one — a cut Huawei is using to position the model for high-volume agent workloads. Developers can call it through Chat or V2 interfaces, with both OpenAI-compatible and Anthropic-compatible endpoints.
On capabilities, Huawei Cloud says the upgrade targets three directions. For agents, the company used more than 100,000 specialized data samples to improve skill triggering, tool selection and tool calling, with claimed gains in deep research tasks and HarmonyOS device control. For code, the model is tuned for HarmonyOS app development in the ArkTS language and for generating operators for Ascend AI chips. For inference, optimized key operators and pipelines deliver much higher throughput in typical deployments — the company says, without giving figures. All of these are vendor statements; Huawei Cloud has not published parameter counts or benchmark results for the 2.1-Pro MaaS version. For reference, openPangu-2.0-Pro, open-sourced in July, was described as a roughly 505B-parameter MoE.
There is a second, quieter story in the release notes: Huawei Cloud is retiring the 2.0 series from MaaS, with openPangu-2.0-Pro and openPangu-2.0-Flash scheduled to go offline on November 30. Notably, the retirement notice does not point users only to Huawei's own replacement — it recommends Zhipu's GLM-5.3 and DeepSeek-V4.1-Flash-1017 as alternatives on the same platform. For a cloud vendor that has spent years building its own foundation model stack, routing outgoing customers toward third-party Chinese open models is a pragmatic acknowledgment of where developer demand actually sits.
Huawei Cloud also plans to fold the model into its own product line, bringing openPangu-2.1-Pro to the AgentArts agent platform, CodeArts coding tools and the OfficeAce office product. That mirrors how the model debuted — embedded in Xiaoyi Work first, exposed as an API second — and keeps the Pangu family anchored to Huawei's HarmonyOS-plus-Ascend ecosystem rather than competing head-on in the general chatbot market.
The strategic picture: China's state-adjacent enterprise cloud is pricing frontier-adjacent models for mass agent deployment, not benchmark bragging rights. A 512K context, sub-4-yuan output pricing and OpenAI-compatible endpoints make openPangu-2.1-Pro an easy default for developers already building on Huawei Cloud — and the fact that the same platform now resells GLM and DeepSeek suggests Huawei sees the MaaS layer, not model exclusivity, as the business it wants to win.
Comments (0)
Log in to join the discussion
Log InNo comments yet