MegaScale (Production)

Closed weights ByteDance,Peking University 530B parameters February 2024

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
ByteDance,Peking University
Organisation type
Industry,Academia
Country
China
Published
23 February 2024
Authors
Ziheng Jiang, Haibin Lin, Yinmin Zhong, Qi Huang, Yangrui Chen, Zhi Zhang, Yanghua Peng, Xiang Li, Cong Xie, Shibiao Nong, Yulu Jia, Sun He, Hongmin Chen, Zhihao Bai, Qi Hou, Shipeng Yan, Ding Zhou, Yiyao Sheng, Zhuo Jiang, Haohan Xu, Haoran Wei, Zhang Zhang, Pengfei Nie, Leqi Zou, Sida Zhao, Liang Xiang, Zherui Liu, Zhe Li, Xiaoying Jia, Jianxi Ye, Xin Jin, Xin Liu

What it does

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

Domain
Language
Task
Language modeling/generation
Approach
Self-supervised learning

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
530B

Production run is stated to have "hundreds of billions of parameters". Since the authors also do a number of experiments with a 530B model, I speculate they've used 530B for the production model.

Training data
tokens

Speculative. Authors note production system was trained on "multi-trillions of tokens". This could refer to training for multiple epochs on the same 300B tokens used to train the 175B and 530B models outlined in more detail in the paper. Alternatively, it could refer to a larger dataset of perhaps 3-9 trillion tokens.

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
3.9 × 10²⁴ FLOP

Speculative. The model is stated to have trained for "several weeks". Assuming 530B parameters and "several" = 3, compute can be estimated from the 175B model's stated PFLOP/sec: 2166.3 aggregate PFlops/sec * 3 weeks * 7 days/week * 24 hours/day * 3600 seconds/hour = 3.9e+24. As an upper bound, say 8e+24.

How it was established
Other

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
NVIDIA A100
Chips used
12,288
Wall-clock time
504 hours (21 days)

Speculative. Authors state "several weeks". For analysis, I've assumed this means around 3 weeks.

Hardware utilisation
MFU 48.0%

Figure 12 shows MFU, which is around 0.484 during stable periods, with short spikes to as low as 0.4. It seems like the overall average is likely minimally affected by the spikes, but I'll round down to 0.48

Power draw
9.7 MW
Compute cost
$2,614,019
Data centre
ByteDance MegaScale training cluster

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

Code for MegaScale (also called veScale) training system are released under Apache Licence: https://github.com/volcengine/vescale The model itself is 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

Improves SOTA in FLOP utilization for distributed LLM training by 1.34X.

Record confidence
Speculative
Citations
302

Sources

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

Reference
MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs
Last updated
25 May 2026

What the numbers mean

Where it came from

MegaScale (Production) was published by ByteDance,Peking University, in China, in February 2024. The organisation is categorised as industry,Academia.

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

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

What went into building it

The training run consumed about 3.9 × 10²⁴ FLOP, on NVIDIA A100. That figure describes the cost of creating it and has no bearing on how quickly it generates text.

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

Answers

MegaScale (Production) — common questions

01

What is MegaScale (Production) used for?

MegaScale (Production) works in Language, and is recorded as handling language modeling/generation. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

02

How much compute was used to train MegaScale (Production)?

Around 3.9 × 10²⁴ FLOP, on NVIDIA A100. 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.

03

What GPU do I need to run MegaScale (Production)?

None. MegaScale (Production) 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.

04

Is MegaScale (Production) open source?

No. MegaScale (Production) has not had its weights published, so it exists only as a service controlled by its owner.

05

How many parameters does MegaScale (Production) have?

MegaScale (Production) has 530B parameters. Production run is stated to have "hundreds of billions of parameters". Since the authors also do a number of experiments with a 530B model, I speculate they've used 530B for the production model. 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.

06

Who created MegaScale (Production)?

MegaScale (Production) was published by ByteDance,Peking University, based in China, categorised as industry,Academia.

07

When was MegaScale (Production) released?

MegaScale (Production) was published in February 2024. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

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.