People visit the Huawei’s Atlas 950 SuperPoD during the World Artificial Intelligence Conference (WAIC) in Shanghai on July 17, 2026. CN-STR/AFP via Getty Images
On July 31, Huawei released the weights, inference code, and technical report for openPangu-2.0-Pro — a 505-billion-parameter language model that completed its full pretraining run on Huawei’s own Ascend 910B NPUs, with no Nvidia hardware involved at any stage. That claim matters because no prior frontier-scale open-weight model — not DeepSeek V4 Pro, not GLM-5.2, not Qwen 3.7 — made it through training without touching a data center loaded with Nvidia A100s or H100s. openPangu-2.0-Pro is the first at 500 billion parameters or above to credibly say otherwise.
The model is now live on GitCode’s Ascend Tribe community and accessible through Huawei Cloud ModelArts Studio. Developers can download the weights and run inference on Ascend hardware today; community efforts to run the model on Nvidia GPUs via standard format conversions are also underway, extending access beyond Huawei’s ecosystem.
What the release proves is real. What it does not prove is that a fully domestic Chinese hardware stack — one independent of foreign fabrication and foreign memory — can yet sustain this kind of training run. That distinction is the story behind the headline.
What "No Nvidia" Actually Means: Supply Chain Behind the Claim
The Ascend 910B NPUs that trained openPangu-2.0-Pro are manufactured primarily on Semiconductor Manufacturing International Corporation’s (SMIC) N+2 process, a 7-nanometer-class node that uses deep ultraviolet (DUV) lithography rather than the extreme ultraviolet (EUV) equipment that TSMC, Samsung, and Intel use for their leading-edge nodes. ASML holds a near-monopoly on EUV systems, and export agreements between the US, the Netherlands, and Japan bar ASML from shipping EUV equipment to China — meaning SMIC is capped at a process node that Nvidia’s Blackwell-class chips surpassed years ago.
But the more immediate supply chain complication is what was inside the Ascend 910B chips themselves at the time they were produced for training runs like openPangu-2.0-Pro. TechInsights teardowns of Ascend 910B and 910C samples found that essentially every chip examined contained dies fabricated at TSMC’s 7nm node — not SMIC’s. Huawei had acquired approximately 2.9 million of those TSMC-fabricated 7nm dies through Sophgo, a Cayman Islands-registered chip design firm, in a scheme US authorities determined violated export controls; TSMC subsequently faced a potential penalty of $1 billion or more under a U.S. Department of Commerce investigation into those transactions. That stockpile of foreign-fabricated chips — the die bank — kept Ascend production volumes viable through 2024 and 2025 while SMIC’s own Ascend production ramped.
As of early 2026, the TSMC die bank is effectively exhausted. Future Ascend production depends entirely on SMIC-fabricated wafers and domestically packaged high-bandwidth memory (HBM). CXMT (ChangXin Memory Technologies), China’s leading DRAM manufacturer, is ramping HBM-class production, but has not yet reached the volume or yield reliability that Samsung or SK Hynix delivers. Samsung HBM stacks were also found in the Ascend 910C teardowns alongside the TSMC compute dies, and China stockpiled approximately 13 million Samsung HBM stacks before export controls on HBM exports tightened.
The upshot is this: openPangu-2.0-Pro demonstrates that the Ascend software and architecture stack can sustain frontier-scale training. It does not yet demonstrate that a fully domestic hardware stack — SMIC-fabricated compute dies paired with CXMT HBM — can do the same, because that is not what trained this model. The next Ascend-based frontier training run will be the real test.
How the Model Was Built: Architecture at 505B Scale
openPangu-2.0-Pro uses a Mixture of Experts (MoE) architecture with 505 billion total parameters, of which approximately 18 billion are activated per token during inference. A routing mechanism selects a subset of specialized expert sub-networks for each input, leaving the rest idle — the design principle that makes large-scale training tractable without paying the full compute cost of a 505-billion-parameter dense model on every query.
Its context window extends to 512,000 tokens, enabling processing of entire codebases, lengthy legal or regulatory documents, or extended multi-session histories without chunking. Pretraining consumed approximately 34 trillion tokens of data.
The attention mechanism is a layered hybrid: Multi-head Latent Attention (MLA) compresses the key-value cache to make 512K-token inference practically feasible, combined with Decoupled Sparse Attention (DSA) and Sliding Window Attention (SWA) layers in a 1:2 ratio. SWA layers handle local-context attention within a defined window; DSA layers capture sparse global dependencies across the full sequence. The 1:2 DSA-to-SWA ratio keeps the quadratic attention cost of long-sequence processing manageable without sacrificing the ability to reason across distant context.
Additional design choices include 4-stream multi-head convolutional residual connections in place of the traditional single residual path — improving signal propagation in very deep networks — and a 3-head Multi-Token Prediction (MTP) module that speculatively predicts multiple tokens per forward pass, raising throughput.
Why the Muon Optimizer Matters for Constrained Hardware
One of the less-discussed but technically significant decisions in openPangu-2.0-Pro’s training is the choice of the Muon optimizer over the Adam-family alternatives that remain standard across most frontier model training runs.
Muon — short for MomentUm Orthogonalized by Newton-Schulz iteration — treats weight matrices as geometric units rather than independent scalars, applying matrix orthogonalization to the momentum buffer before each update step. Research comparing Muon and AdamW at scale has found that Muon achieves substantially greater computational efficiency, reaching equivalent validation loss levels at significantly reduced token budgets — an advantage equivalent to having materially more compute available. On a constrained Ascend cluster where per-chip throughput is lower than on Nvidia’s equivalent hardware, that efficiency is not a marginal advantage. It is the mechanism that makes this training run viable at this scale on this hardware.
DeepSeek V4 Pro, the 1.6-trillion-parameter model released in April 2026, also uses Muon for most of its training modules. Kimi K2 and GLM-5.2 have adopted variants of Muon as well. The optimizer is becoming the de facto standard for large-scale MoE training runs on Ascend-class hardware, precisely because its efficiency advantage offsets a portion of the per-chip throughput deficit.
Post-Training: The OPD Approach to Capability Fusion
openPangu-2.0-Pro’s post-training pipeline addresses one of the hardest problems in frontier model development: how to combine capabilities across multiple domains without the catastrophic forgetting that typically degrades earlier capabilities when a model is fine-tuned on new tasks.
The pipeline ran in three stages. Supervised Fine-Tuning (SFT) integrated what the technical report describes as "fast and slow thinking" styles — aligning the model’s responses with both quick pattern-completion tasks and deliberate multi-step reasoning. Multiple specialist Reinforcement Learning (RL) stages then separately targeted distinct capability domains, training specialist-model variants each optimized for a specific domain without the interference that would occur if all domains trained simultaneously.
The third stage — Online Policy Distillation (OPD) — is the architecturally novel step. Rather than applying the specialist variants sequentially (which would restore catastrophic forgetting) or averaging their weights (which loses the specialization), OPD runs online during training, fusing capabilities from the specialist variants back into a single general-purpose model while all weights are live. The result is a model that carries the performance improvements from each specialist domain without losing the others.
No prior open-source model at this parameter count has documented this three-stage OPD pipeline in its technical report. Whether the technique is genuinely novel at frontier scale or an application of known distillation principles to a new architectural context will become clearer as independent researchers read the technical report now available via GitCode.
Performance: What Is and Is Not Yet Known
Huawei claims openPangu-2.0-Pro achieves roughly double the single-card throughput on Ascend hardware compared with other major open-source models of comparable scale. That figure is Huawei’s own. No independent benchmark platform — Artificial Analysis, BenchLM, or comparable — had published openPangu-2.0-Pro results as of this article’s publication.
The Flash variant — 92 billion total parameters, 6 billion activated — released on June 30 provides partial evidence. Community testing of Flash broadly confirmed competitive performance on Chinese- and English-language tasks. The Flash model’s architecture has already been merged into ik_llama.cpp via community contributions, enabling inference on non-Ascend hardware. A capability gap relative to frontier models like DeepSeek V4 Pro has been estimated at roughly one to one-and-a-half generations, though that estimate comes from architectural comparison rather than systematic benchmarking.
For context: the Council on Foreign Relations’ December 2025 analysis found that the best current US AI chips are approximately five times more powerful than Huawei’s best offerings in total processing performance, with that gap projected to widen to 17 times by 2027. Per-chip constraints do not directly translate to model quality — Muon’s efficiency gains and MoE’s inference economy can partially offset hardware gaps — but the constraint remains real and independent benchmarks will reveal its magnitude.
The RAND Corporation documented that iFlytek, a major Chinese AI company, experienced a three-month delay in model development when it switched from Nvidia hardware to Huawei Ascend 910B chips for training. That was before Muon adoption at scale; it is not clear whether the same delay applies to a team already optimizing natively for Ascend, as Huawei’s own research team was.
Who the Model Is For — and Who Can Actually Use It
Three audiences have immediate reasons to pay attention.
Researchers in export-restricted or Nvidia-restricted regions. Countries and institutions that cannot legally or practically access Nvidia hardware have faced a growing gap between the models they can train locally and the ones the wider AI research community builds on. openPangu-2.0-Pro is the first open-weight model at 500B+ parameters where the entire training stack — chip, interconnect, framework, model — can in principle be reproduced domestically using non-US-origin components, provided those components include Ascend-class NPUs and sufficient HBM supply. Over 100 countries have signed the Bangkok Declaration committing to AI sovereignty programs; openPangu-2.0-Pro is now the most technically detailed reference for what an Ascend-native frontier training run looks like. Malaysia’s 2025 announcement of an Ascend-based sovereign AI initiative underscores that this reference architecture has real buyers.
AI/ML researchers. The combination of MLA, DSA/SWA hierarchical attention, OPD-based post-training, and Muon at 500B+ scale has not appeared in open literature at this parameter count before. The technical report is now publicly available. Independent verification of the architecture choices and their performance tradeoffs will take weeks; the community evaluation period is just beginning.
Infrastructure operators and cloud providers. The model is available through Huawei Cloud ModelArts as a hosted inference service and as downloadable weights for self-hosted deployment on Ascend hardware. Community ports to Nvidia GPU inference via standard conversion formats are in progress. The 512K context window is among the longest available in any open-weight model, which may distinguish it for long-document processing pipelines regardless of the hardware story.
The Legal Conditions Attached to Any Huawei Model
No article about an AI model from a Chinese company is complete without naming the legal conditions that apply regardless of open-source licensing, corporate structure, or server location.
China’s National Intelligence Law (2017), Article 7, requires that all Chinese organizations and citizens "support, assist, and cooperate with national intelligence work in accordance with the law." This obligation applies to Huawei and any affiliated community that operates under Chinese law, including the openJiuwen and Ascend Tribe communities that host and maintain openPangu-2.0-Pro. The Cybersecurity Law (2016, amended January 2026), Article 28, further requires network operators to provide "technical support and assistance" to public security and national security agencies on demand. The Data Security Law (2021) adds cross-border data transfer restrictions and additional government access provisions.
These are fixed legal conditions, not contested claims or speculative risks. They apply regardless of: the openPangu License’s permissive terms, the open-source availability of the weights, the physical location of the inference server, and any privacy policy Huawei publishes. A developer who downloads the weights and self-hosts the model on infrastructure entirely outside China is insulated from API-based data access. A developer who uses Huawei Cloud ModelArts for hosted inference is not — the data processed by that inference passes through infrastructure operated by a company subject to the laws above.
This disclosure is not a reason to avoid the model. It is information a reader needs before choosing a deployment model.
Where the Ecosystem Gap Still Sits
The openPangu-2.0-Pro release is an architectural and strategic milestone. It is not an ecosystem parity claim.
Nvidia’s CUDA platform, launched in 2007, has over 4 million registered developers and decades of embedded library optimization. Huawei’s CANN (Compute Architecture for Neural Networks) platform is actively developing, but developers consistently report that porting CUDA-optimized production code to CANN introduces friction, performance regressions, and debugging overhead that CUDA workflows do not require. Huawei’s commitment to open-source CANN and MindSpore by year-end 2026 is a meaningful step; it is also a public acknowledgment that ecosystem maturity is the most significant remaining competitive deficit.
The model’s single-card throughput claims are self-reported and unverified by independent auditors. The inference throughput gap between Ascend and Nvidia hardware at comparable task loads has been measured at approximately 17–19 tokens per second versus 25 or more on Nvidia-backed alternatives — a gap that does not prevent deployment but does affect application design, latency budgets, and user experience in interactive contexts.
Nvidia’s CEO Jensen Huang has said that US export policy has "largely backfired" and given Nvidia "zero percent" market share in China. openPangu-2.0-Pro is the strongest single data point in support of that assessment. But a 500B-parameter training demonstration on a hardware stack whose supply chain was partially supported by pre-restriction foreign components is not the same as a demonstration that a fully domestic stack can replicate the result without degradation. That demonstration has not yet happened.
Frequently Asked Questions
Can Huawei really train a frontier AI model without Nvidia chips?
Yes, as openPangu-2.0-Pro demonstrates. The model completed its full pretraining run on Ascend 910B NPUs with no Nvidia A100 or H100 hardware in the training cluster. The important nuance is that the Ascend 910B chips used in that training run were themselves partially fabricated using TSMC-manufactured dies and foreign-sourced Samsung HBM memory — supply chain components whose availability is now under tighter export controls. The model proves that the Ascend software and architecture stack can sustain 500B+ parameter training; it does not yet prove that a fully domestic hardware supply chain can do the same.
Is openPangu-2.0-Pro as capable as DeepSeek V4 Pro or Claude Fable?
Independent benchmarks were not available at publication — the model released July 31, and systematic evaluations from platforms like Artificial Analysis require days to weeks after weights become public. Architectural comparison suggests openPangu-2.0-Pro, with 18 billion active parameters from a 505-billion-parameter MoE pool and a 512K token context window, should perform comparably to dense models in the 50–70-billion-parameter class on single-turn tasks, while benefiting from the larger expert pool on knowledge-intensive tasks. A capability gap of roughly one to one-and-a-half generations relative to DeepSeek V4 Pro has been estimated, but that estimate is architectural inference, not measured benchmark data.
Who should actually consider deploying openPangu-2.0-Pro?
Three categories: sovereign AI programs in jurisdictions that cannot access Nvidia hardware and need a frontier-scale training blueprint; AI researchers who want to study the OPD post-training architecture or MLA/DSA attention design; and infrastructure operators already running on Ascend hardware who can take advantage of the 512K context window for long-document processing. Developers seeking the best available performance per token for production inference workloads should wait for independent benchmarks. For most production use cases today, DeepSeek V4 Pro or GLM-5.2 remain better-benchmarked options.
What are the privacy risks of using openPangu-2.0-Pro through Huawei Cloud?
Downloading the weights and self-hosting the model on infrastructure outside China carries no data-access obligation to the Chinese government — once downloaded, the weights are yours to run. Using Huawei Cloud ModelArts for hosted inference routes your data through infrastructure operated by Huawei, a company subject to China’s National Intelligence Law (2017), Article 7, which legally requires all Chinese organizations to cooperate with government intelligence requests on demand, and the Cybersecurity Law’s Article 28, which requires technical assistance to public security agencies. These are fixed legal conditions, not contractual risks. Readers making deployment decisions for sensitive workloads should choose self-hosted inference on non-Huawei-Cloud infrastructure.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.