RetNet
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
- Microsoft Research,Tsinghua University
- Organisation type
- Industry,Academia
- Country
- United States of America, China
- Published
- 17 July 2023
- Authors
- Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, Furu Wei
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling
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
- 6.7B
- Training data
- 100,000,000,000 tokens
- Batch size
- 4,000,000
Table 2 They later mention testing the memory and throughput of a 13B-parameter model, but it doesn't sound like they trained it long enough to test its perplexity.
4M
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 × 10²¹ FLOP
- How it was established
- Operation counting
C = 6ND = 6 * 6.7 billion * 100 billion "We train the models with 512 AMD MI200 GPUs." Another rough method: Table 4 shows throughput is 8642.2 words per second during training on an 8xA100 setup, with batch size of 8192. This suggests 2543.1 hours to go through the full 100B tokens ~= 75B words. (8 * 3.12e14) * (2543.1 * 3600) * (0.3) = 6.9e21 FLOPs (low confidence in this as actual training setup used different hardware and much larger batch size, but same OOM is encouraging)
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
MIT for code looks like model code not train code: https://github.com/microsoft/unilm/tree/master/retnet
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Record confidence
- Confident
- Citations
- 646
Sources
Where this record came from and when it was last checked.
- Reference
- Retentive Network: A Successor to Transformer for Large Language Models
- Last updated
- 25 May 2026
What the numbers mean
Where it came from
RetNet was published by Microsoft Research,Tsinghua University, in the country recorded as United States of America, during July 2023. The category the publisher falls under is industry,Academia.
It works in the domain of Language, and is recorded as performing the task of language modeling.
This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.
How it was trained
Producing it required arithmetic totalling around 4 × 10²¹ FLOP. That figure measures what producing the model cost, and has no bearing on how fast it answers.
The training set ran to roughly 100,000,000,000 tokens of text.
Answers
RetNet — common questions
RetNet— when was it released?
It was published in July 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.
RetNet— what is it used for?
It works in the domain of Language, and is recorded as handling the task of language modeling. 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.
RetNet— how much compute was used to train it?
Training consumed around 4 × 10²¹ FLOP. 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.
RetNet— what GPU do I need to run it?
None. This 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.
RetNet— is it open source?
No. Its weights have not been published, so it exists only as a service controlled by its owner.
RetNet— how many parameters does it have?
It has a parameter count of 6.7B. Table 2 They later mention testing the memory and throughput of a 13B-parameter model, but it doesn't sound like they trained it long enough to test its perplexity. 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.
RetNet— who created it?
It was published by Microsoft Research,Tsinghua University, based in United States of America, an organisation categorised as industry,Academia.
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.