NLM

Closed weights Carnegie Mellon University (CMU),University of California San Diego 247M parameters September 2021

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
Carnegie Mellon University (CMU),University of California San Diego
Organisation type
Academia,Academia
Country
United States of America
Published
9 September 2021
Authors
Junxian He, Graham Neubig, Taylor Berg-Kirkpatrick

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
247M

The core language model is semi-parametric. There is a non-parametric component which involves learning a datastore which maps the training data to context vectors. This paper improves on the kNN-LM model primarily by reducing the size of the datastore to improve throughput. They also add a small 512 parameter neural net to adaptively adjust the interpolation hyperparameter. They use the pre-trained model from https://arxiv.org/abs/1809.10853, which had 247M trainable parameters, so the model i…

Training data
tokens

103M training tokens in WikiText-103

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
2.8 × 10¹⁹ FLOP

The base model used was 247M parameters and trained for 286k updates with an effective batch size of 65,536 tokens. 6 * 247M * 286k * 65536 = 2.778e19 FLOPs Generating the datastore involves a single forward pass over all of the training examples, for another 2 * 247M * 103M = 5.088e16 FLOPs They also train the small retrieval adapter, but this converges in "several minutes with a single GPU" and so is unlikely to contribute significantly. The only other computationally expensive step is the…

How it was established
Operation counting

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 GeForce RTX 3090
Chips used
1
Power draw
388 W

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
Open source

MIT for code: https://github.com/jxhe/efficient-knnlm

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
123
Benchmark data
NLM

Sources

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

Reference
Efficient Nearest Neighbor Language Models
Last updated
25 May 2026

What the numbers mean

About this model

NLM was published by Carnegie Mellon University (CMU),University of California San Diego, in United States of America, in September 2021. It comes out of academia,Academia.

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

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

How it was trained

Producing it required around 2.8 × 10¹⁹ FLOP of arithmetic, on NVIDIA GeForce RTX 3090, which is a statement about the training budget rather than about inference.

Answers

NLM — common questions

01

What is NLM used for?

NLM works in Language, and is recorded as handling 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.

02

How much compute was used to train NLM?

Around 2.8 × 10¹⁹ FLOP, on NVIDIA GeForce RTX 3090. 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 NLM?

None. NLM 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 NLM open source?

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

05

How many parameters does NLM have?

NLM has 247M parameters. The core language model is semi-parametric. There is a non-parametric component which involves learning a datastore which maps the training data to context vectors. This paper improves on the kNN-LM model primarily by reducing the size of the datastore to improve throughput. They also add a small 512 parameter neural net to adaptively adjust the interpolation hyperparameter. They use the pre-trained model from https://arxiv.org/abs/1809.10853, which had 247M trainable parameters, so the model in this paper one has 247M + 512 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.

06

Who created NLM?

NLM was published by Carnegie Mellon University (CMU),University of California San Diego, based in United States of America, categorised as academia,Academia.

07

When was NLM released?

NLM was published in September 2021. 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.