Research

Huawei Pangu and the Ascend Stack 2026

Huawei Pangu model history + Ascend chip stack: Pangu 5.5 (Jun 2025), openPangu open-source (Jun 2025), Pangu Ultra MoE 718B on 6,000 Ascend NPUs, Ascend 910C 600K-unit target 2026.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: May 2026

What this is

Huawei's Pangu line is the only Chinese frontier-class LLM trained entirely on Chinese-stack hardware (Ascend NPUs). The combined Pangu + Ascend story is the vertical-integration case study for China's AI independence. This page is a 2026-05-15 reference.

Pangu Release Timeline (2024-2026)

DateReleaseNotable
2024Pangu 5.0 seriesFoundation models on Ascend
2025 (HDC 2025)Pangu Models 5.5NLP / CV / multimodal / prediction / scientific computing
2025Pangu Ultra (dense)135B params, 128K context, 8,192 Ascend NPUs, 13.2T tokens
May 2025Pangu Ultra MoE718B parameters, 6,000+ Ascend NPUs
Jun 30, 2025openPangu open-source7B + 72B Pro MoE released openly
Oct 30, 2025Pangu LM 8.6.0Service overview release

Ascend Chip Stack (2026)

ChipStatusNotable
Ascend 910C600,000 units targeted in 2026 (2x 2025)Flagship training chip; competes with H20 / B-series
Ascend 950PRQ1 2026 launchCard + SuperPoD server formats
Ascend total 2026 die outputUp to 1.6M diesAcross model variants
Approval statusOn China government-approved AI hardware listCo-listed with Cambricon; NVIDIA excluded

Six Things the Pangu + Ascend Story Tells You

  1. Pangu is the only frontier LLM trained exclusively on Chinese-stack hardware. Every other Chinese frontier model (DeepSeek, Qwen, Kimi, GLM) still uses or has used NVIDIA at some training stage.
  2. openPangu (Jun 2025) was Huawei's open-source response. 7B + 72B Pro MoE released openly; positions Huawei alongside DeepSeek and Qwen on developer mindshare.
  3. Ascend 910C output doubling in 2026 to ~600K units is the supply signal. Combined with Ascend 950PR card/SuperPoD launch, total die output could reach 1.6M.
  4. Day-0 adaptation of DeepSeek-V4 on Ascend (alongside Cambricon and Hygon) marked the transition from "lagging" to "simultaneous" Chinese hardware support.
  5. Pangu Ultra MoE at 718B parameters trained on 6,000+ Ascend NPUs is the largest Chinese-hardware-only training run published to date.
  6. Government-approved supplier status gives Huawei structural advantage in Chinese state / SOE deployments where NVIDIA is now excluded.

What This Means for AI Visibility

Pangu-served surfaces are mostly inside Huawei Cloud and Chinese SOE / state deployments. Most global brands have limited Pangu exposure today, but the vertical-stack story matters because it sets the trajectory for Chinese government and regulated-sector AI deployments. Brands targeting Chinese enterprise customers should expect Pangu to appear in tenders and deployment shortlists alongside DeepSeek and Qwen.

Methodology

Pangu details from Wikipedia Huawei PanGu, Developing Telecoms on Pangu 5.5, MarkTechPost on Pangu Ultra MoE, and Huawei Cloud Pangu product page. Ascend chip data from Tom's Hardware on Ascend ecosystem and Bloomberg on Ascend output doubling.

How Presenc AI Helps

Presenc AI tracks brand visibility on Huawei Pangu surfaces (Pangu LM API, openPangu deployments). For brands targeting Chinese state, telecom, financial-services, and energy customers — where Huawei + Ascend deployments are concentrated — Pangu coverage is an increasingly important channel.

Frequently Asked Questions

Partially. openPangu (released June 30, 2025) includes a 7B model and a 72B Pro MoE variant under an open licence. Pangu Ultra (135B) and Pangu Ultra MoE (718B) remain proprietary, deployed primarily via Huawei Cloud.
Vertical integration. Huawei is the only Chinese vendor with both frontier-class models and frontier-class training hardware (Ascend). For Chinese government, state-owned enterprise, telecom, and regulated-sector buyers, the Huawei stack is often the default. The benchmark gap to GPT-5 / Claude Opus matters less in that segment than supply security.
For Chinese deployments under export controls, increasingly yes. Ascend 910C output is doubling to ~600K units in 2026; the chip is now on China's government-approved AI hardware list while NVIDIA is excluded. Performance per watt still trails NVIDIA's latest generation, but the supply security and government approval status outweigh the efficiency gap inside China.
Huawei Cloud (primary), openPangu via HuggingFace for the 7B and 72B Pro MoE open releases, and on-prem Ascend deployments for enterprise customers. The proprietary tier (Pangu Ultra, Pangu Ultra MoE) is deployed via Huawei Cloud Model-as-a-Service offerings.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.