Huawei dropped a surprise on July 31: the full weights, inference code, and a detailed technical report for openPangu-2.0-Pro, a 505-billion-parameter mixture-of-experts language model trained entirely on the company’s own Ascend 910B accelerators—no Nvidia A100 or H100 hardware in sight. It’s the first time a model at this scale has credibly bypassed Nvidia’s ecosystem for training. But as soon as you dig past the headline, a more complicated hardware story emerges: the chips used in that training run almost certainly contained foreign-made silicon from TSMC and memory from Samsung, stockpiled before export controls tightened. That doesn’t diminish the engineering feat, but it does clarify what “trained without Nvidia” actually means—and what it doesn’t.

Inside the openPangu-2.0-Pro Release

The model appeared first on GitCode’s Ascend Tribe community and is also accessible through Huawei Cloud ModelArts Studio. Developers can download the weights and run inference on Ascend hardware today, while community efforts to convert the model for Nvidia GPUs are already underway. A companion technical report lays out the architecture and training methodology.

At its core, openPangu-2.0-Pro is a massive MoE design: 505 billion total parameters, but only about 18 billion are activated per token. A routing mechanism picks a small set of specialized expert sub-networks for each input, leaving the rest idle. This approach dramatically cuts inference cost compared to a dense 505B model, and it also made the training run feasible on Ascend hardware.

Other specs that matter: a 512,000-token context window that can swallow entire codebases or regulatory filers without chunking, and a pretraining corpus of roughly 34 trillion tokens. The attention scheme is layered—Multi-head Latent Attention compresses the key-value cache, combined with Decoupled Sparse Attention and Sliding Window Attention layers in a 1:2 ratio. That hybrid approach keeps long-sequence processing tractable. Additional details include multi-head convolutional residual connections and a multi-token prediction module that speculatively predicts several tokens per forward pass.

One of the more interesting decisions was the choice of optimizer. Instead of the common AdamW, Huawei used Muon (MomentUm Orthogonalized by Newton-Schulz iteration), which treats weight matrices as geometric units and applies orthogonalization to the momentum buffer. Research suggests Muon can achieve equivalent validation loss with fewer training tokens—a critical advantage when you’re running on accelerators that don’t match Nvidia’s per-chip throughput. Post-training followed a three-stage pipeline: supervised fine-tuning, specialist reinforcement learning for distinct capability domains, and Online Policy Distillation to fuse those capabilities back into a single general model without catastrophic forgetting.

Huawei’s own throughput numbers claim roughly double the single-card performance on Ascend compared to other open-source models of similar scale. Independent benchmarks from platforms like Artificial Analysis have not yet appeared, as the model was only just released.

Who Stands to Benefit—and How

For most everyday Windows users, openPangu-2.0-Pro won’t show up on your desktop. But it signals an accelerating competition in AI infrastructure—and that competition will ultimately filter down into smarter assistants, better code generation tools, and more capable enterprise software. If you’re just watching the AI race, this is a proof point that Nvidia alternatives can now train frontier-scale models.

For developers and power users, the immediate benefit is access to a 500B-parameter playground. The weights are available, and community ports to non-Ascend hardware are being built. The Flash variant (92B total, 6B active) has already been merged into ik_llama.cpp, so running the Pro model on a high‑end workstation is not out of the question—though you’ll need a serious GPU memory setup and likely quantization. A word of caution: early conversions may introduce performance regressions, especially around the custom attention layers. If you just want to experiment without the hardware headache, Huawei Cloud ModelArts offers hosted inference, but that brings legal considerations (more on that below).

IT professionals and system administrators should pay attention for three reasons. First, the 512K context window can handle very long documents, making the model interesting for processing legal contracts, technical manuals, or extended support logs. Second, if your organization operates in a jurisdiction that cannot legally access Nvidia hardware, openPangu-2.0-Pro documents an Ascend-based training blueprint that can be reproduced locally—provided you can source the accelerators and memory. Third, the legal environment matters: if you use Huawei Cloud for inference, your data passes through infrastructure governed by China’s National Intelligence Law (Article 7) and Cybersecurity Law (Article 28), which legally require operators to assist government intelligence and public security agencies. Self-hosting the weights on your own infrastructure, anywhere in the world, avoids that obligation.

The Trail of Chips: Export Controls and the Ascend Buildup

Understanding today’s announcement requires rewinding to September 2020, when TSMC stopped supplying Huawei under U.S. export controls. At that point, Huawei’s chip roadmap hit a wall. But the company found a workaround: it acquired roughly 2.9 million TSMC-fabricated 7nm dies through Sophgo, a Cayman Islands-registered design firm. Those dies wound up inside Ascend 910B and later 910C products, as TechInsights teardowns confirmed. Meanwhile, Huawei stockpiled Samsung high‑bandwidth memory (HBM) before further restrictions kicked in. In short, the Ascend chips that trained openPangu-2.0-Pro were not fully domestic; they relied on a pre‑restriction inventory of foreign-made components.

With that die bank now largely exhausted, future Ascend production depends on SMIC-fabricated wafers and Chinese HBM from companies like CXMT. CXMT is ramping but hasn’t yet matched Samsung or SK Hynix volumes or reliability. So when Huawei says “trained without Nvidia,” the claim is accurate at the cluster level, but the chips themselves were built on borrowed silicon.

This doesn’t diminish the software achievement. Getting a 505B MoE model through pretraining on Ascend required maturing Huawei’s CANN stack, MindSpore framework, networking, and kernel libraries to a point where they could sustain such a run. The RAND Corporation documented that a major Chinese AI company experienced a three-month delay when switching from Nvidia to Ascend 910B for training—and that was before Muon became widely adopted. The optimizer’s efficiency gain likely made the difference between a viable run and an infeasible one.

OpenPangu-2.0-Pro arrives in a year that has already seen DeepSeek V4 Pro (1.6T params, also using Muon) and Kimi K2 push the boundaries of Chinese AI. What sets Huawei’s model apart is the public release of a complete Ascend-native training recipe at this scale.

Your Next Move

If you’re simply curious, wait. Independent benchmarks will land in weeks, giving a clearer picture of where the model sits against DeepSeek, GLM, and Qwen. Until then, any performance claims are Huawei’s own.

For developers: grab the weights from GitCode if you’re prepared for a heavy engineering task. Community conversion efforts are worth tracking, but don’t expect drop‑in production performance on Nvidia GPUs yet. Use ModelArts for quick tests only with non‑sensitive data, given the legal environment. Read the technical report—the OPD post-training and attention architecture are novel enough that they could inspire other projects, even if you never deploy the model itself.

For IT decision-makers: if you already run Ascend hardware, this model is a ready-made candidate for long‑context document analysis. Yet Ascend’s ecosystem still trails CUDA’s maturity by years; porting existing CUDA‑optimized code will be painful. Huawei has committed to open‑sourcing CANN and MindSpore by end of 2026, which could narrow that gap, but it isn’t here yet. If data sovereignty is a concern, self‑host the weights on non‑Chinese infrastructure—that cuts the legal tether to Huawei Cloud. No special licensing restrictions appear to block commercial use, though you should verify with the openPangu License.

The Road to Full Autonomy

The real milestone hasn’t happened yet. A comparable training run on SMIC‑fabricated compute dies paired with CXMT‑supplied HBM would prove a fully domestic Chinese AI stack. Until that model ships, openPangu-2.0-Pro is a brilliant software‑and‑systems demo, not a declaration of semiconductor independence. Watch for SMIC’s 7nm‑class yield improvements and CXMT’s HBM ramp. Also monitor whether future Huawei models adopt the next‑gen Ascend architecture or stick with the existing 910 series. For now, the release is a reminder that AI training is no longer synonymous with Nvidia—but the hardware supply chain still is.