DensePhrases

Open weights Korea University,Princeton University December 2020

No estimate

No hardware requirements for this model

This model's weights are open, but no parameter count has been published for it. Every memory and speed figure starts from that number, so we would rather show nothing than a fabricated estimate.

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
Korea University,Princeton University
Organisation type
Academia,Academia
Country
Korea (Republic of), United States of America
Published
23 December 2020
Authors
Jinhyuk Lee, Mujeen Sung, Jaewoo Kang, Danqi Chen

What it does

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

Domain
Language
Task
Question answering

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.

Training data
58,000,000 tokens

from appendix D "The number of generated questions is 327,302 and 1,126,354 for SQuAD and Natural Questions, respectively." assuming 40 words per question we get around ~ 58M

Epochs
4

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
2.1 × 10¹⁸ FLOP

flops = (8) * (1215 * 10**10) * (20 * 3600) * 3 // 10 = 2099520000000000000 (num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate) model of GPU from appendix B (Titan Xp) number of GPUs from table in appendix A flops from https://www.techpowerup.com/gpu-specs/titan-xp.c2948

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
NVIDIA TITAN Xp
Chips used
8
Chip-hours
160
Wall-clock time
20 hours

appendix A row 3

Power draw
4.1 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
Open — downloadable
Model access
Open weights (unrestricted)
Training code
Open source

Apache 2.0: https://github.com/princeton-nlp/DensePhrases

How it is classified

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

Why it is tracked
SOTA improvement

from abstract "our model DensePhrases improves over previous phrase retrieval models by 15%-25% absolute accuracy and matches the performance of state-of-the-art retriever-reader models. " "we achieve state-of-the-art performance on two slot filling tasks (Petroni et al., 2021), using less than 5% of the training data."

Record confidence
Speculative
Citations
128

Sources

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

Reference
Learning Dense Representations of Phrases at Scale
Last updated
25 May 2026

What the numbers mean

What this model is

DensePhrases was published by Korea University,Princeton University, in Korea (Republic of), in December 2020. It comes out of academia,Academia.

It works in Language, and is recorded as doing question answering.

The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached.

Training and provenance

Training it took roughly 2.1 × 10¹⁸ FLOP of computation, on NVIDIA TITAN Xp — a measure of what producing the model cost, not of how fast it answers.

The training set ran to roughly 58,000,000 tokens.

Its inclusion criterion is sOTA improvement.

Answers

DensePhrases — common questions

01

What is DensePhrases used for?

DensePhrases works in Language, and is recorded as handling question answering. These are the areas it was designed around; they describe intent rather than a hard boundary.

02

Where can I download DensePhrases?

The weights for DensePhrases are published, though we do not hold a repository link for it. This site calculates hardware requirements rather than hosting model files.

03

How much compute was used to train DensePhrases?

Around 2.1 × 10¹⁸ FLOP, on NVIDIA TITAN Xp. 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.

04

What GPU do I need to run DensePhrases?

We cannot say. DensePhrases has open weights, but no parameter count has been published for it, and every memory and speed calculation starts from that number. We would rather show nothing than a fabricated estimate.

05

Is DensePhrases open source?

Its weights are published, so DensePhrases can be downloaded and run on your own hardware. Note that open weights is not the same as open source in the full sense — it says nothing about the training data, the training code, or the commercial terms attached.

06

How many parameters does DensePhrases have?

No parameter count has been published for DensePhrases, which is why no memory or speed figure appears on this page.

07

Who created DensePhrases?

DensePhrases was published by Korea University,Princeton University, based in Korea (Republic of), categorised as academia,Academia.

08

When was DensePhrases released?

DensePhrases was published in December 2020. 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.