No Nvidia Required: Huawei Ships 505B Open-Weight AI Model

On July 31, Huawei released the weights, inference code, and technical report for openPangu-2.0-Pro — a 505-billion-parameter Mixture-of-Experts language model that completed its full pretraining run on Huawei's own Ascend 910B NPUs. No Nvidia A100. No H100. No CUDA. The model is now live on GitCode's Ascend Tribe community and mirrored on Hugging Face, with hosted inference available through Huawei Cloud ModelArts Studio.

This is not a research paper. It is downloadable weights from a model that trained 34 trillion tokens across approximately 505 billion parameters — 18 billion activated per token — with a 512K-token context window. Every forward pass, backward pass, and optimizer step ran on Ascend's CANN runtime rather than Nvidia's CUDA, on an interconnect Huawei designed itself. openPangu-2.0-Pro is the first open-weight model at 500 billion parameters or above that credibly claims to have never touched an Nvidia GPU.

What Happened

The release follows a choreographed sequence. At HDC 2026 in June, Huawei executive Richard Yu announced openPangu 2.0 would open-source seven major components — pretraining code, post-training code, training operators, weights, and inference code — starting June 30. The 92-billion-parameter openPangu-2.0-Flash shipped first, followed by the full Pro variant on July 31. Additional components, including the complete training stack, are scheduled to roll out through the second half of 2026.

The architecture is aggressively modern. openPangu-2.0-Pro uses Multi-head Latent Attention with a hybrid DSA+SWA layer pattern at a 1:2 ratio — sparse global aggregation layers interleaved with sliding-window local attention layers, reducing the compute and memory cost of long-context inference. The standard residual topology is upgraded to a four-branch mHC architecture that increases representational diversity. A three-head Multi-Token Prediction module speculatively generates three additional tokens per forward pass, accelerating inference throughput.

The training optimization choice is where the Ascend-native story gets concrete. Instead of AdamW — the industry default — Huawei used Muon, a momentum-based optimizer that treats weight matrices as geometric units, applying matrix orthogonalization to the momentum buffer before each update step. Research comparing Muon and AdamW at scale has found Muon achieves equivalent validation loss at substantially reduced token budgets. On constrained Ascend clusters where per-chip throughput is lower than Nvidia equivalents, that efficiency advantage is not marginal. It is the mechanism that makes a 505B-parameter training run viable on this hardware.

Why It Matters

Until this release, every frontier-scale open-weight model — DeepSeek V4 Pro, GLM-5.2, Qwen 3.7, Kimi K3 — trained on data centers loaded with Nvidia A100s or H100s, even when the companies behind them faced export restrictions. The distinction between "trained by a Chinese lab" and "trained on Nvidia hardware" collapsed because they were the same thing.

openPangu-2.0-Pro breaks that collapse. The model card explicitly states Ascend-native training, the inference code targets Ascend's omni-infer framework, and the architecture choices — Muon over AdamW, hybrid attention over pure dense — reflect hardware-aware decisions consistent with Ascend's performance profile. No independent auditor has verified the training run, and the claim should be treated as credible but not independently confirmed. But the weights exist, the architecture is documented, and the benchmarks — AIME 2026 at 95.4 in thinking mode, GPQA-Diamond at 87.9, SWE-bench Verified at 68.5, LiveCodeBench V6 at 85.7 — put it in competitive territory with Western frontier models from 12 months ago.

The geopolitical dimension is inescapable. Over 100 countries have signed the Bangkok Declaration committing to AI sovereignty programs. For nations and institutions that cannot legally or practically access Nvidia hardware, openPangu-2.0-Pro is now the most technically detailed reference architecture for what an Ascend-native frontier training run looks like. Malaysia's 2025 Ascend-based sovereign AI initiative already has real buyers for this reference design.

The Supply Chain Asterisk

Here is what the "no Nvidia" claim does not mean: it does not mean no foreign silicon touched the training cluster.

TechInsights teardowns of Ascend 910B and 910C samples found that essentially every chip examined contained compute dies fabricated at TSMC's 7nm node — not SMIC's. Huawei acquired approximately 2.9 million of those TSMC-fabricated 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. Samsung HBM memory stacks were also found in the Ascend 910C teardowns, and China stockpiled approximately 13 million Samsung HBM stacks before export controls tightened.

That die bank — the TSMC-fabricated silicon inside chips labeled "Made in China" — kept Ascend production volumes viable through 2024 and 2025. 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 from CXMT, China's leading DRAM manufacturer. CXMT is ramping HBM-class production but has not yet reached the volume or yield reliability Samsung or SK Hynix delivers.

openPangu-2.0-Pro demonstrates the Ascend software and architecture stack can sustain frontier-scale training. It does not yet demonstrate a fully domestic hardware stack can do the same, because that is not what trained this model. The next Ascend-based frontier training run — the one that runs on SMIC compute dies paired with CXMT HBM — will be the real test of hardware sovereignty.

Builder Impact

For developers and infrastructure teams, openPangu-2.0-Pro matters in three concrete ways.

First, if you own or are procuring Ascend hardware, the deployment path is direct: pull weights, use the omni-infer framework, and run inference with Huawei's optimized operators. The 512K context window is among the longest available in any open-weight model, which distinguishes it for long-document processing pipelines regardless of the hardware story.

Second, if you are on Nvidia GPUs, budget real engineering time. Community ports via standard format conversions are in progress, but the inference code is Ascend-native and optimized operators will need reimplementation for CUDA. Start with the 92B Flash variant, which is more manageable for single-node deployment and already has broader community support.

Third, the license contains a landmine. Huawei distributes the model under its custom OpenPangu Model License Agreement 2.0, which explicitly bars access, deployment, or use within the European Union. If your application has EU users or EU-hosted infrastructure, openPangu-2.0-Pro is not an option regardless of your hardware stack. This is not a standard open-source license — it is a strategic restriction embedded in the distribution terms, and it effectively bifurcates the open-weight ecosystem along geopolitical lines.

What Comes Next

The sequence Huawei has committed to matters more than the model itself. If the remaining components — pretraining code, post-training code, and training operators — ship on schedule through the end of 2026, the Ascend ecosystem will have, for the first time, a complete open-source stack from silicon to inference. That is what CUDA had for years and what kept Nvidia's moat intact: not just the chips, but the software ecosystem that made switching expensive.

Nvidia CEO Jensen Huang has said US export policy "largely backfired" and gave Nvidia "zero percent" market share in China. openPangu-2.0-Pro is the strongest single data point in support of that assessment. But the hardware supply chain asterisk remains — the model is proof of software and architecture readiness, not domestic silicon readiness. The next training run, and the one after that, will determine whether Huang's assessment becomes permanent or proves premature.

For AI builders watching from outside the US-China tech contest, the lesson is simpler: the assumption that frontier AI requires Nvidia hardware has just been published as falsifiable. Download the weights and verify.