Reported 15 Nov 2026 · Labs A one-minute read
Huawei's new open-source flagship arrived with no weights, no parameter count and a price list.
Huawei Cloud put openPangu-2.1-Pro on its model service on November 15, calling it open source. Caveat: it published no weights, parameters or technical report, and stopped selling its genuinely open 2.0 models.
Open-source Pangu meant downloadable weights and a published technical report.
It can mean a metered API on Huawei's own cloud.
Why it matters
The 2.0 models were real open weights, 92 billion and 505 billion parameters, with training code. Huawei is retiring them and telling customers to migrate to DeepSeek and GLM models instead. What carries the open-source name now runs only on Huawei's cloud, tuned for HarmonyOS and Ascend chips.
Established
- Huawei Cloud listed openPangu-2.1-Pro on its model service on November 15.
- Upgrades were claimed in agent development, AI coding and inference.
- It is sold per token, with cached input priced far lower.
Not established
- That weights never follow. Huawei released earlier Pangu weights in stages.
- Its size or context length. Huawei has published neither.
- How it performs. No benchmark or technical report accompanied the launch.
The five-minute readThe record, the timeline, the weight and 6 sources
The record
Reported, not verified. Huawei Cloud's own announcement sits on huaweicloud.com, which refused us, and Huawei Cloud's international newsroom carries nothing. The record rests on IT Home, Smzdm, Sina Finance and TokenPost, all quoting Huawei Cloud and all agreeing on the date and the three upgrade areas. Reading Huawei's own page, or a weights release, by December 16 moves it.
Huawei Cloud's model-as-a-service platform listed openPangu-2.1-Pro on Sunday November 15, 2026. TokenPost filed at 9:31 am Eastern, 14:31 Universal Time. IT Home filed at 21:16 Beijing time, 13:16 Universal Time. Sina Finance and PANews carried the same notice the same day. All four attribute it to Huawei Cloud and quote the same line, that the MaaS platform has officially brought openPangu-2.1-Pro online. Three upgrade areas are named: agent development, AI coding, and inference optimization. IT Home reports that Skills triggering, tool selection and tool invocation were tuned on more than 100,000 specialized data entries, and that the coding work targets HarmonyOS app development in ArkTS and Ascend operator generation. Named use cases are deep research, HarmonyOS system control, HarmonyOS app development, Ascend operator generation and inference pipeline optimization. The model is to be wired into Huawei Cloud's AgentArts, CodeArts and OfficeAce products. IT Home gives the price in yuan per million tokens: 0.8 for input, 3.2 for output, 0.016 for cached input. No parameter count, active parameter count, context length, architecture, benchmark score, technical report, licence or weights download appears in any account. The release was pre-announced. TokenPost reported on November 6 that Huawei Cloud planned the November 15 listing and had notified existing users by text message. Smzdm sets out the predecessors by contrast: openPangu-2.0-Flash at 92 billion total and 6 billion active parameters, and openPangu-2.0-Pro at 505 billion total and 18 billion active, both open-sourced in mid-2026 with weights and technical reports, with pretraining, supervised fine-tuning and reinforcement learning code published on September 28. Smzdm reports Huawei Cloud stopped selling the 2.0 models on November 6 and takes them offline on November 30, 2026 at 00:00, and advised migrating Flash users to DeepSeek-V4.1-Flash-1017 and Pro users to GLM-5.3. Smzdm also notes openPangu-2.0-Pro scored 65.7% on BrowseComp against 91.2% for Kimi K3 and 83.4% for DeepSeek-V4-Pro, under conditions it says were not uniform, and that Huawei claimed up to twice the single-card inference throughput of mainstream open models on Ascend hardware.
How it got here
- Mid-2026
Huawei open-sources openPangu-2.0-Flash and openPangu-2.0-Pro, 92 billion and 505 billion total parameters, with weights and technical reports. Open-source Pangu means a download.
- Sep 28
Huawei publishes the pretraining, supervised fine-tuning and reinforcement learning code for the openPangu line, completing a staged release.
- Nov 5
IT Home finds openPangu-2.1-Pro already running inside Huawei's Xiaoyi Work app, marketed as HarmonyOS task enhancement, with no specifications disclosed.
- Nov 6
Huawei Cloud notifies existing users by text message that openPangu-2.1-Pro lists on November 15. Smzdm reports the 2.0 models stopped being sold the same day, and go offline on November 30.
- Nov 15
openPangu-2.1-Pro goes live on Huawei Cloud MaaS, priced per token, with no weights, parameter count, context length, benchmark or technical report published.
Huawei's HarmonyOS and Ascend developers, who gain a tuned model and lose a downloadable one; the 2.0 users told to migrate to DeepSeek and GLM, now running rivals' models on Huawei's cloud; GitCode, where the weights would land; and anyone counting Pangu as an open model.
What to watch
- Would strengthen itHuawei publishes no openPangu-2.1-Pro weights, parameter count or technical report by March 31, 2027, and the 2.0 models leave Huawei Cloud as scheduled on November 30, 2026.
- Would strengthen itHuawei publishes openPangu-2.1-Pro weights on GitCode or Hugging Face, with a parameter count and a technical report, by March 31, 2027.
- Would strengthen itHuawei names an Ascend-only or HarmonyOS-only licence term for openPangu-2.1-Pro, or reports paying customers for it, by June 30, 2027.
Why it weighs this much
Cost is medium. Training and serving a flagship is expensive, and retiring two open-weight models weeks after announcing their replacement throws away whatever the downloads bought. Commitment is medium: a price list and a dated migration notice are public and awkward to reverse, but a product listing is not a contract. Reversibility is easy, and Huawei just proved it by delisting the 2.0 models it had open-sourced earlier this year. Persistence is a year. The Ascend and HarmonyOS tuning points at a stack strategy that outlasts any single model, while the model itself will be replaced.
The words
- Open weights
- A model whose trained parameters can be downloaded and run on your own hardware. Without them, open source describes the paperwork rather than anything you can hold.
- Model as a service
- Renting a model through an interface priced per unit of text, instead of running it yourself. The provider keeps the weights.
- Ascend operator
- A low-level routine that makes a model run fast on Huawei's own AI chips. Generating these is the bottleneck in leaving Nvidia.
- Cached input
- Text the provider has already processed and stored, charged at a small fraction of fresh text.
Sources
- IIT Home: openPangu-2.1-Pro lists on Huawei Cloud, quoting Huawei Cloud's notice, with per-token pricing, Nov 15, 2026ithome.com
- TTokenPost: Huawei Cloud adds openPangu-2.1-Pro to its MaaS platform, Nov 15, 2026, 9:31 am Easterntokenpost.com
- TTokenPost: Huawei Cloud plans a Nov 15 release of openPangu-2.1-Pro, with text-message notice to users, Nov 6, 2026tokenpost.com
- PSmzdm: the new Pangu flagship that publishes no parameters, and the 2.0 models withdrawn from sale, Nov 15, 2026post.smzdm.com
- FSina Finance: open-source Pangu model openPangu-2.1-Pro officially live on Huawei Cloud, Nov 15, 2026finance.sina.cn
- MIT Home: Xiaoyi Work carries openPangu-2.1-Pro first, no specifications disclosed, Nov 5, 2026m.ithome.com
Your call
Which explanation do you believe?
A call is final. Other people's calls show after yours.