Pangu Ultra

Closed weights Huawei 135B parameters April 2025

No estimate

No hardware requirements for this model

The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation
Huawei
Organisation type
Industry
Country
China
Published
10 April 2025
Authors
Yichun Yin, Wenyong Huang, Kaikai Song, Yehui Tang, Xueyu Wu, Wei Guo, Peng Guo, Yaoyuan Wang, Xiaojun Meng, Yasheng Wang, Dong Li, Can Chen, Dandan Tu, Yin Li, Fisher Yu, Ruiming Tang, Yunhe Wang, Baojun Wang, Bin Wang, Bo Wang, Boxiao Liu, Changzheng Zhang, Duyu Tang, Fei Mi, Hui Jin, Jiansheng Wei, Jiarui Qin, Jinpeng Li, Jun Zhao, Liqun Deng, Lin Li, Minghui Xu, Naifu Zhang, Nianzu Zheng, Qian…

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
Language
Task
Code generation, Language modeling/generation

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters
135B
Training data
13,200,000,000,000 tokens

13.2 trillion tokens

Batch size
8,388,608

2048 * 4096

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute
1.1 × 10²⁵ FLOP

When compared to Llama 3.1 405B, Pangu Ultra achieves better scores on most of the challenging benchmarks, while utilizing only about 29% of the training FLOPs required by Llama 405B. Compute = 6 FLOP/token/param * 135e9 params *13.2e12 tokens = 1.069200e+25 FLOP This is consistent with 29% of Llama 405B's compute: 3.8e25*0.29=1.1e25.

How it was established
Operation counting,Comparison with other models

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware
Huawei Ascend 910B
Chips used
8,192
Hardware utilisation
MFU 52.0%

"Through careful tuning of load balancing across PP and VPP stages, we are able to achieve approximately 43% MFU on an 8,192 NPU cluster as a baseline." ... "Overall, we achieve over 52% Model FLOPs Utilization (MFU) when training Pangu Ultra on 8,192 Ascend NPUs."

Power draw
6.4 MW

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights
Closed — provider access only
Model access
Hosted access (no API)
Training code
Unreleased

"Our model and system will be available for our commercial customers"

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Likely above 10²³ FLOP
Yes
Why it is tracked
SOTA improvement

A closer examination reveals that Pangu Ultra excels on Chinese benchmarks, surpassing both Qwen 2.5 72B and DeepSeek V3, the current best-performing Chinese model. In addition, when compared to Llama 3.1 405B, Pangu Ultra achieves better scores on most of the challenging benchmarks, while utilizing only about 29% of the training FLOPs required by Llama 405B. These results suggest the effectiveness of our model architecture and the high quality of our training data. Table 3, Tabke 4: seems that…

Record confidence
Confident
Citations
16

Sources

Where this record came from and when it was last checked.

Reference
Pangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUs
Last updated
25 May 2026

What the numbers mean

What this model is

Pangu Ultra was published by Huawei, in China, in April 2025. It comes out of industry.

It works in Language, and is recorded as doing code generation, Language modeling/generation.

Its weights were never published, so it can only be reached through its provider. No graphics card changes that.

How it was trained

Training it took roughly 1.1 × 10²⁵ FLOP of computation, on Huawei Ascend 910B — a measure of what producing the model cost, not of how fast it answers.

Around 13,200,000,000,000 tokens went into training it.

The reason it appears in this catalogue at all is sOTA improvement.

Answers

Pangu Ultra — common questions

01

Who created Pangu Ultra?

Pangu Ultra was published by Huawei, based in China, categorised as industry.

02

When was Pangu Ultra released?

Pangu Ultra was published in April 2025.

03

What is Pangu Ultra used for?

Pangu Ultra works in Language, and is recorded as handling code generation, Language modeling/generation. These are the areas it was designed around; they describe intent rather than a hard boundary.

04

How much compute was used to train Pangu Ultra?

Around 1.1 × 10²⁵ FLOP, on Huawei Ascend 910B. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

05

What GPU do I need to run Pangu Ultra?

None. Pangu Ultra is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.

06

Is Pangu Ultra open source?

No. Pangu Ultra has not had its weights published, so it exists only as a service controlled by its owner.

07

How many parameters does Pangu Ultra have?

Pangu Ultra has 135B parameters. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

Source

Original publication

Record last updated 25 May 2026

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.