Huawei Open-Sourced a 505B Model. Europeans Aren't Allowed to Run It.

TL;DR

On July 31 Huawei open-sourced openPangu-2.0-Pro: a mixture-of-experts model with roughly 505B total parameters, about 18B active per token, a 512K context window, and around 34T tokens of pretraining. Weights, inference code, and a technical report all landed at once, and the model card says the whole thing was trained on Ascend NPUs. Two details matter more than the parameter count. Section 3.1 of the license forbids using the model anywhere in the European Union, and on the 22 benchmarks that appear on both Huawei model cards, the 92B openPangu-2.0-Flash sibling beats the 505B flagship on nine of them.

What actually shipped

The GitCode drop and the Hugging Face mirror carry the same card. Pro is 505B total with 18B activated, 512K context, and a thinking / non-thinking split baked into the same checkpoint. Hugging Face tags the repo at 541B params in BF16, which puts the full download a little over a terabyte on disk.

The architecture notes are the interesting part, because Huawei is not copying anyone's homework here. Attention is MLA combined with a 1:2 layer ratio of DSA and SWA, so most layers only do local sliding-window work while a minority handle sparse global aggregation. Think of it as a reader who skims paragraph by paragraph and is only occasionally allowed to flip back anywhere in the book, which is how you make a 512K window affordable instead of quadratic.

On top of that: a 4-stream mHC residual topology instead of the standard single residual stream, three MTP heads for speculative decoding, and the Muon optimizer in pretraining. Post-training combines SFT, several rounds of specialized RL, and what Huawei calls online distillation for capability fusion.

The numbers Huawei published

In thinking mode Pro reports 95.4 on AIME 2026 (Avg@16), 87.9 on GPQA-Diamond (Avg@4), 85.7 on LiveCodeBench V6 (Avg@3), and 68.5 on SWE-bench Verified (Avg@3). With a Python tool in the loop, AIME goes to 97.9.

The card is refreshingly willing to publish the ugly rows too. Humanity's Last Exam comes in at 27.1, CL-Bench at 19.7, and Apex at 17.7 in thinking mode and 2.1 without it. Nobody is saturating the hard evals, and Huawei did not quietly leave them out.

Standard caveat, and it is not a small one: these are self-reported numbers with no independent replication yet, on a model most people cannot easily run. Treat the table as a claim, not a result.

The sibling problem

Here is the part worth staring at. Flash is 92B total and 6B active, roughly 5.5x smaller on paper and 3x smaller in active compute, and it was trained on the same 34T tokens. Line up the 22 benchmarks that appear in thinking mode on both cards and Flash wins nine of them.

The wins are not all rounding errors either. Flash takes TAU2-Bench 88.0 to 81.1, Multichallenge 68.4 to 62.9, IFEval 95.9 to 94.5, DeepCodeBench 76.5 to 74.2, and PinchBench 85.6 to 84.0. It also edges SysBench, AgentIF, CL-Bench, and AIME with Python. On LiveCodeBench V6 the 505B model wins by six tenths of a point.

Pro earns its keep on the things you would expect scale to buy: GPQA-Diamond by 4.2, IMO-AnswerBench by 7.8, BBEH by 7.2, BrowseComp by 8.7, SWE-bench Verified by 5.4. And in non-thinking mode the gap is brutal in Pro's favor on code, 74.5 to 50.9 on LiveCodeBench V6, which suggests the small model leans much harder on its reasoning budget to stay in the fight.

The honest read is that this is what a well-distilled small sibling looks like in 2026, not that the flagship is broken. It is still a slightly awkward launch note when your headline model has to share a page with the cheap one.

Then you read the license

The model ships under the OPENPANGU MODEL LICENSE AGREEMENT VERSION 2.0. It is perpetual, worldwide, non-exclusive, no-charge, royalty-free, and commercial use is fine. It is also, per Section 3.1, considerably less worldwide than that sentence implies.

You represent and warrant that You will not, access, download, install, run, deploy, integrate, modify, or otherwise use the Model, directly or indirectly, within the European Union.

Section 4.2 adds an attribution requirement: any product or service you build on it, including another AI model, has to display "Powered by openPangu" and a Huawei trademark notice. Section 3.2 is a patent retaliation clause, and Huawei reserves the right to update the agreement at any time, at which point you either accept the new terms or stop using the model.

To be clear, the EU carve-out is not new to this release. It has been in the openPangu family license for a while, and it plausibly reads as Huawei declining to take on GPAI-provider obligations rather than a statement about Europe. Fine. But it is the reason you should not file this next to an OSI-approved license in your head, and if your compliance team has an "open weights are fine" rule, this is the model that breaks it.

Can you actually run it?

Realistically, not on your desk. Just over a terabyte of BF16 weights puts this outside homelab range, and Huawei's own deployment path points at openPangu-2.0-Infer, an inference project built around the omni-infer framework targeting Ascend 910C. There is no advertised day-one vLLM or llama.cpp story, which is a meaningful difference from how a Qwen or DeepSeek drop lands.

That is the strategic point of the whole exercise. Huawei is not really competing for your 4090. It is demonstrating that a frontier-scale training run, the silicon, the interconnect, and the serving stack can all be domestic, and shipping the weights is the proof-of-work. The Flash sibling, at 92B total and 6B active, is the one that actually has a shot at running on hardware you own.

Key Takeaways

• openPangu-2.0-Pro is real and downloadable: 505B total, 18B active, 512K context, ~34T pretraining tokens, weights plus inference code plus a technical report, released July 31.
• It was trained on Ascend NPUs. The card names Huawei's own silicon and mentions no other hardware, which is the actual headline for anyone tracking whether the Nvidia-free training path is credible at this scale.
• The 92B Flash sibling wins 9 of 22 shared benchmarks, including TAU2-Bench by 6.9 points, at roughly a fifth of the total parameters.
• Section 3.1 bans use in the European Union outright: no accessing, downloading, running, or deploying, directly or indirectly.
• Section 4.2 requires a "Powered by openPangu" badge on anything you build with it, including derived models.
• All benchmark numbers are self-reported with no third-party replication yet, and the practical serving path is Ascend-first via omni-infer.

Sources: openPangu-2.0-Pro model card, openPangu-2.0-Flash model card, Ascend Tribe openPangu-2.0-Pro repository, OPENPANGU MODEL LICENSE AGREEMENT VERSION 2.0, AIbase release report, openPangu-2.0-Infer