PanGu-Σ

Closed weights Huawei Noah's Ark Lab 1.1T parameters March 2023

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 Noah's Ark Lab
Organisation type
Industry
Country
China
Published
20 March 2023
Authors
Xiaozhe Ren, Pingyi Zhou, Xinfan Meng, Xinjing Huang, Yadao Wang, Weichao Wang, Pengfei Li, Xiaoda Zhang, Alexander Podolskiy, Grigory Arshinov, Andrey Bout, Irina Piontkovskaya, Jiansheng Wei, Xin Jiang, Teng Su, Qun Liu, Jun Yao

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, Translation, Question answering
Approach
Self-supervised learning
Numerical format
FP16

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
1.1T

"In this work, we present PanGu-Σ , a large language model with sparse architecture containing 1.085 trillion parameters."

Training data
329,000,000,000 tokens

329B tokens ~= 247B words

Epochs
1
Batch size
524,288

"We train PanGu-Σ with global batch size of 512 with sequence length of 1024 for each sample"

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
4.7 × 10²³ FLOP

It has sparse architecture, so we can't use C=6ND. "We develop PanGu-Σ model under the framework of MindSpore and train it on a cluster with only 512 Ascend 910 AI Accelerators with 329 billion tokens over 100 days." 100 days * 512 processors * 320 teraFLOPS/processor * 33% utilization = 4.67e+23 FLOP https://www.wolframalpha.com/input?i=100+days+*+512+*+320+terahertz+*+0.33

How it was established
Hardware

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 910
Chips used
512
Chip-hours
1,228,800
Wall-clock time
2,400 hours (100 days)

We develop PanGu-Σ model under the framework of MindSpore 5 and train it on a cluster with only 512 Ascend 910 AI Accelerators [28] with 329 billion tokens over 100 days.

Power draw
316.5 kW

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
Unreleased
Training code
Unreleased

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

"Our experimental findings show that PanGu-{\Sigma} provides state-of-the-art performance in zero-shot learning of various Chinese NLP downstream tasks." "The results on WMT20 translation task. PanGu-Σ outperforms previous Chinese-English SOTA pre-trained large model with a large margin." " The PanGu-Σ outperforms the current state-of-the-art model PanGu-Coder by 1.4 point on the pass@1 for MBPP tasks" "Table 5: Zero-shot results of Chinese downstream tasks. Compared to ERNIE 3.0 Titan, PanGu…

Record confidence
Confident
Citations
78

Sources

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

Reference
PanGu-Σ: Towards Trillion Parameter Language Model with Sparse Heterogeneous Computing
Last updated
6 January 2026

What the numbers mean

About this model

PanGu-Σ was published by Huawei Noah's Ark Lab, in China, in March 2023. It comes out of industry.

It works in Language, and is recorded as doing code generation, Language modeling, Translation, Question answering.

Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.

Training and provenance

Producing it required around 4.7 × 10²³ FLOP of arithmetic, on Huawei Ascend 910, which is a statement about the training budget rather than about inference.

It was trained on about 329,000,000,000 tokens of text.

It is tracked in the underlying dataset for one reason in particular: sOTA improvement.

Answers

PanGu-Σ — common questions

01

Who created PanGu-Σ?

PanGu-Σ was published by Huawei Noah's Ark Lab, based in China, categorised as industry.

02

When was PanGu-Σ released?

PanGu-Σ was published in March 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

03

What is PanGu-Σ used for?

PanGu-Σ works in Language, and is recorded as handling code generation, Language modeling, Translation, Question answering. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

04

How much compute was used to train PanGu-Σ?

Around 4.7 × 10²³ FLOP, on Huawei Ascend 910. 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-Σ?

None. PanGu-Σ 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-Σ open source?

No. PanGu-Σ has not had its weights published, so it exists only as a service controlled by its owner.

07

How many parameters does PanGu-Σ have?

PanGu-Σ has 1.1T parameters. "In this work, we present PanGu-Σ , a large language model with sparse architecture containing 1.085 trillion 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 6 January 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.