Huawei Trained a 505B Model Without Touching One Nvidia Chip

On July 31, Huawei put a 505-billion-parameter AI model up for anyone to download. The weights, the inference code, and a full technical report, all public. That happens fairly often now. What has never happened before is the other part: according to Huawei, the entire training run, all 34 trillion tokens of it, ran on its own Ascend chips. No Nvidia hardware anywhere in the loop.

Every big model you have ever used, GPT, Gemini, Claude, DeepSeek, Kimi, exists because of Nvidia GPUs. openPangu-2.0-Pro is the first model at this scale that claims otherwise.

What actually shipped

Per the model card on Hugging Face and Huawei's own announcement, openPangu-2.0-Pro is a mixture-of-experts model with about 505 billion total parameters, roughly 18 billion active per token, a 512K context window, and around 34 trillion tokens of pretraining data. For comparison, per RuntimeWire, DeepSeek-V3 has a 128K context window, and so do Qwen's and Kimi's big releases. Huawei is claiming four times that.

This was not a surprise drop. Huawei announced a seven-part open-source plan at its developer conference in June, released a smaller 92 billion parameter sibling called Flash on June 30, and said the Pro model was coming in July. It landed on the last day of the month, per AI2Work's writeup.

One honest caveat before the specs impress anyone: every benchmark number published so far, 68.5 on SWE-bench Verified, 95.4 on AIME 2026 for the thinking version, comes from Huawei itself. No independent lab has evaluated this model yet. And the "zero Nvidia" claim, while consistent with everything in the technical report, has not been audited by anyone outside Huawei. AI2Work also notes that DeepSeek reportedly went back to Nvidia chips after running into trouble training on Ascend, which tells you how hard this is supposed to be.

The number that matters is zero

The reason this release is a big deal has nothing to do with benchmarks. It is about export controls.

The US has spent years trying to keep advanced Nvidia chips out of China. The current state of that fight, per The Register and Bloomberg: Washington approved limited H200 sales to China in January under a rule that takes a 25% cut of revenue, roughly ten Chinese companies got US clearance, and Beijing never approved the imports on its side. Nvidia's own guidance now counts its China data center revenue as zero. Per StockWireX, China was about 17% of Nvidia's total revenue in fiscal 2024.

Meanwhile Huawei has been building the replacement. Its Ascend 910C chip delivers roughly 80% of an H100's performance on paper, per AI2Work, made on SMIC's 7nm process instead of TSMC's leading edge. Bloomberg has reported Huawei is targeting around 600,000 of those chips this year, double last year. openPangu-2.0-Pro is the proof of concept that the whole stack, chips, interconnect, training software, model, works end to end with nothing American in it that sanctions can switch off.

The fine print nobody is headlining

Here is where it gets interesting if you actually read the license. I did.

The weights come under Huawei's custom OpenPangu Model License 2.0, not a standard open-source license like Apache or MIT. Two clauses stand out. Section 3.1 says you may not access, download, run, or deploy the model "directly or indirectly, within the European Union." An entire continent, written out of the license.

Section 4.2 says any product built on the model has to display "Powered by openPangu" plus a Huawei trademark notice. And Section 10 says Huawei can change the terms at any time and you have to comply or stop using it. The inference code itself is Apache 2.0, per RuntimeWire, but the model is not.

There is also a hardware catch. The official inference framework, omni-infer, targets Ascend chips. If you want to run this on Nvidia hardware, per AI2Work, you are waiting on community ports. And at roughly 1.08 terabytes of weight files, per RuntimeWire, nobody is running this at home regardless. Even the parameter count has a quirk: Huawei says 505 billion, Hugging Face's metadata says 541 billion, and the repo does not explain the gap.

I run open-weight Chinese models every day, Kimi K3 is the daily brain of my agent setup, so I checked what it would take to try this one. The answer, for me and for basically any individual developer, is nothing. There is no hardware path. This model is built for companies and governments that need a frontier-scale model no export-control regime can revoke, and the license terms make clear Huawei gets to decide who that includes.

Where this leaves things

What is verifiable today: a 505 billion parameter model with public weights, a 512K context window, and a license that excludes the EU now exists, and Huawei says no Nvidia chip touched its training. What is not verifiable yet: whether it performs anywhere near its self-reported numbers, and whether the training claim survives an independent look.

The last time the assumption was that Chinese labs could not compete without American chips, DeepSeek spent a week as the biggest story in tech. That was about doing more with fewer Nvidia chips. This is the first serious claim of doing it with none.