gpt-neo-2.7B

2.72B parameters, text in, text out, 2K tokens of context.

Verified 2026-10-25 Rank #367 in LLMs · score 80.6 method v1 Sources 3 Downloads 30d 16.7k · #471 of 619 tracked

Can I run it?1.6GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?MITPermissivelicence file2026-10-24

Do people use it?16.7kdownloads, 30 days#471 of 619 we track · below the median−1 likes, downloads −3% this week503 likesHugging Face2026-10-24

Worth reading

Picked by rule from 15 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 561 HN pts

GPT Neo: open-source GPT model, with pretrained 1.3B & 2.7B weight models

The GPT-Neo repository offers an archival implementation of model parallel GPT-3-style models using the mesh-tensorflow library, superseded by GPT-NeoX for GPU efficiency.

Training and inference is officially supported on TPU and should work on GPU as well.
Google colab provides tpu-v8s for free, which should be enough to finetune our models up to GPT3XL (1.5B parameter) sizes.

github.com 2021-03-21 561 HN pts · 127 comments 3k words about the gpt neo family family

Latest · 4 years ago

H3 – Outperforming GPT-Neo-2.7B with only 2 attention layers

twitter.com 2023-01-23 10 HN pts

15 best of 15 tracked

The 2 picks above, plus these 13. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleGPT-Neo – Building a GPT-3-sized model, open source and freeeleuther.ai · 2021-01-18 · 725 HN pts · official
ArticleFew-Shot Learning in Practice: GPT-Neo & 'HuggingFace' Accelerated Inference APIhuggingface.co · 2021-06-04 · 69 HN pts
ArticleUsing GPT-J and GPT-Neo, GPT-3 open-source alternatives, with few-shot learningnlpcloud.io · 2021-11-10 · 24 HN pts
CodeHugging Face Transformers – Big Model Inference, Bloom, GPT Neo-X, LongT5 etc.github.com · 2022-06-16 · 2 HN pts
ArticleGPT-Neo vs. GPT-3: Are Commercialized NLP Models That Much Better?medium.com · 2021-04-09 · 2 HN pts
ArticleGPT-Neo on a M1 (macOS 12)lazycoder.ro · 2022-02-03 · 5 HN pts
ArticleGPT-3's free alternative GPT-Neo is something to be excited aboutventurebeat.com · 2021-05-15 · 5 HN pts
ArticleGPT-3 open-source alternatives: GPT-Neo and GPT-Jnlpcloud.io · 2021-12-14 · 4 HN pts
ArticleNovelAI, a new AI storyteller being built on GPT-Neonovelai.medium.com · 2021-05-31
ArticlePretrain GPT-Neo for Open Source GitHub Copilot Modeldiscuss.huggingface.co · 2021-07-03
ArticleHow to monitor the performance of a fine-tuned GPT-Neo Model in Productionblog.cerebrium.ai · 2023-02-28
DiscussionFew-shot learning NLP with GPT-J and GPT-Neonews.ycombinator.com · 2021-07-22
DiscussionHow to run inference with GPT-Neo for freetwitter.com · 2021-04-05

Show all 13

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 1 quantized builds and 100 Spaces.

As a service0no API provider listed
On own hardware1quantized builds
Built on3725 fine-tunes · 12 adapters · 0 merges
In apps100+Hugging Face Spaces

Top Spaces running it: KBaba7/Quant ♥ 51 · EleutherAI/magma ♥ 46 · lxe/simple-llm-finetuner ♥ 35 · devendergarg14/Paraphrasing_with_GPT_Neo ♥ 26

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 1.6 GB at INT4, 3.3 GB at FP8, 6.5 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2022-03-02), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

38 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

38 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

10 rows

Benchmarks 10 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (10)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported Commonsense QA sets · Chat and follow instructions	72.14% percent	GPT-Neo 2.7B	model author	2026-10-07
author-reported HellaSwag · Chat and follow instructions	42.73% percent	GPT-Neo 2.7B	model author	2026-10-07
author-reported Lambada Acc	62.22% percent	GPT-Neo 2.7B	model author	2026-10-07
author-reported Lambada PPL	5.63 other	GPT-Neo 2.7B	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	24.72% percent	GPT-Neo 2.7B	model author	2026-10-07
author-reported Pile BPB	0.72 other	GPT-Neo 2.7B	model author	2026-10-07
author-reported Pile PPL	5.65 other	GPT-Neo 2.7B	model author	2026-10-07
author-reported PubMedQA	57.54% percent	GPT-Neo 2.7B	model author	2026-10-07
author-reported Wikitext PPL	11.39 other	GPT-Neo 2.7B	model author	2026-10-07
author-reported Winogrande · Chat and follow instructions	56.50% percent	GPT-Neo 2.7B	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

The model uses a GPTNeoForCausalLM architecture and was trained on the Pile dataset.^[1]^[2]

architectures
“trained on the Pile”

Generated 2026-10-07 15:09 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

3 use cases

What it is good at 3 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Chat and follow instructionsauthor-reportedAuthor reports Winogrande 56.50%Author reports Hellaswag 42.73%Author reports Piqa 72.14%
Reason through hard problemsauthor-reportedAuthor reports MathQA 24.72%
Run on a laptop or phonestructuraldecision-model estimate p=0.86 fits: single-gpu

Facts

20 fields

Listed in LLMs

Facts

gpt-neo-2.7B has 2.72B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 2,048 tokens (2K). (config.json max_position_embeddings, observed 2026-10-24)
License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
Created 2022-03-02, last modified 2023-07-09. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	GPTNeoForCausalLM
Model type	gpt_neo
Params by dtype	F32: 2,651,307,520 · U8: 67,108,864
Size band	M
MoE	no
Context length	2,048 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2101.00027
License	MIT (mit)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for gpt-neo-2.7B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	6.5 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	3.3 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)	1.6 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

38 variants

Lineage

Original model Family: EleutherAI/gpt-neo 38 derivatives on the Hub

finetune: 25quantized: 1adapter: 12

Top variants (38 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
OptimalScale/gpt-neo2.7B-inst-tuning	finetune	—	—	—	mit	OptimalScale	0	7.9k
pszemraj/gpt-peter-2.7B	finetune	—	2.79B	M	—	pszemraj	2	785
subhrokomol/gpt-neo-1.3B-hindi	finetune	—	127M	S	apache-2.0	subhrokomol	0	880
a-g-2000/gpt-neo-2.7B-tiny_skspr_finetuned	finetune	—	2.65B	M	mit	a-g-2000	1	482
YALCINKAYA/opsgenius_s	quantized	—	2.65B	M	apache-2.0	YALCINKAYA	0	323
KimByeongSu/gpt-neo-2.7B-cs-finetuning-7	finetune	—	2.65B	M	mit	KimByeongSu	0	244
wakaflocka17/gptneo-imdb-finetuned	finetune	—	—	—	apache-2.0	wakaflocka17	0	241
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning	finetune	—	—	—	mit	KimByeongSu	0	208
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning_MAGNET_valid	finetune	—	—	—	mit	KimByeongSu	0	176
Striker-7/finetune	adapter	—	—	—	mit	Striker-7	0	167
mia-llm/gpt-neo-2.7B-wikitext2raw-LoRA-ilhan	adapter	—	—	—	—	mia-llm	0	125
Maarij-Aqeel/Mental_Health_chatbot	adapter	—	—	—	—	Maarij-Aqeel	0	111
xyzhu1225/ImBD	finetune	—	—	—	apache-2.0	xyzhu1225	2	0
mia-llm/gpt-neo-2.7B-xsum-LoRA-ilhan	adapter	—	—	—	—	mia-llm	0	75
KimByeongSu/gpt-neo-2.7B-cs-finetuning-7-60000	finetune	—	2.65B	M	mit	KimByeongSu	0	67
KimByeongSu/gpt-neo-2.7B-cs-finetuning-filtered-7	finetune	—	2.65B	M	mit	KimByeongSu	0	64
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning_MAGNET	finetune	—	—	—	mit	KimByeongSu	0	63
mia-llm/gpt-neo-2.7B-AGnews-LoRA-ilhan	adapter	—	—	—	—	mia-llm	0	63
KimByeongSu/gpt-neo-2.7B-cs-finetuning-500-1	finetune	—	2.65B	M	mit	KimByeongSu	0	57
muhammadjasim12/results	finetune	—	2.65B	M	mit	muhammadjasim12	0	57
KimByeongSu/gpt-neo-2.7B-lama	finetune	—	2.65B	M	mit	KimByeongSu	0	53
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning_aeda	finetune	—	—	—	mit	KimByeongSu	0	49
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning_eda	finetune	—	—	—	mit	KimByeongSu	0	45
KimByeongSu/neo-2.7B-fine-same	finetune	—	2.65B	M	mit	KimByeongSu	0	36
Rabadios/fine_tuned_gpt_neo	finetune	—	2.65B	M	mit	Rabadios	0	35
adalib/torchrec-data-gpt-neo-2.7B-prefix	adapter	—	—	—	—	adalib	0	32
wakaflocka17/ensemble-majority-voting-imdb	finetune	—	—	—	apache-2.0	wakaflocka17	0	31
scholl99/GPT_NEO_PROMPT_TUNING_CAUSAL_LM	adapter	—	—	—	—	scholl99	0	29
baebee/GPTNeoX-mychatgptconvos-adapter	adapter	—	—	—	—	baebee	0	27
adalib/torchdata-data-gpt-neo-2.7B-prefix	adapter	—	—	—	—	adalib	0	27
interneuronai/az-gptneo	adapter	—	—	—	—	interneuronai	0	26
sedaghat989Dev/floria-lora-2.0	adapter	—	—	—	—	sedaghat989Dev	0	26
adalib/sqlmodel-data-gpt-neo-2.7B-prefix	adapter	—	—	—	—	adalib	0	24
KimByeongSu/gpt-neo-2.7B_LAMA_TREx_finetuning_MAGNET_same	finetune	—	—	—	mit	KimByeongSu	0	23
Omkar7/gpt-neo-2.7B-english-quotes-finetuned	finetune	—	—	—	mit	Omkar7	0	0
Litzy619/V0410MP3_neo	finetune	—	2.68B	M	mit	Litzy619	0	0
2point0/RadenGpt	finetune	—	—	—	other	2point0	0	0
xyzhu1225/ImBD-inference	finetune	—	—	—	apache-2.0	xyzhu1225	0	0

Repository changelog

23 commits

Repository changelog 23 commits

Date	Commit	SHA
2023-07-09	Adding `safetensors` variant of this model (#12)	e24fa29
2023-06-21	Update README.md	8542cf1
2022-12-08	Correcting metadata	a0e5677
2021-12-31	Update README.md	51568a6
2021-12-31	Updated tags to correctly link with the Pile	5e755b1
2021-09-11	Update README.md	6f23148
2021-09-11	Update README.md	88f8889
2021-07-04	add flax model	0b8087b
2021-05-21	Update README.md	b41a392
2021-05-18	Updated citation info	1172dff
2021-05-06	Updated LFS tracked files	df3bd66
2021-05-05	Addition of Rust model	9b4ecbc
2021-04-12	Update README.md	3e1c92c
2021-04-06	Update model max length	5640038
2021-04-04	Update README.md	9130025
2021-03-31	Update README.md	058e8e2
2021-03-31	Update README.md	40fb054
2021-03-31	Update config.json	f5e76ba
2021-03-30	Sample	5957e80
2021-03-30	Sample	7dc5d8d
2021-03-30	Add model card	5813d1e
2021-03-30	add files	ef4a5f2
2021-03-30	initial commit	d31f415

Articles and coverage

15 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

GPT-Neo – Building a GPT-3-sized model, open source and free

eleuther.ai2021-01-18officialHN front pageabout the gpt neo familyHN 725 pts · 252 comments

Few-Shot Learning in Practice: GPT-Neo & 'HuggingFace' Accelerated Inference API

huggingface.co2021-06-04authorityabout the gpt neo familyHN 69 pts · 3 comments1,201 words

GPT-Neo enables few-shot learning for NLP tasks by requiring only 3-4 examples to achieve good results, as it is about 60x smaller than GPT-3. The 🤗 Accelerated Inference API hosts this 2.7B parameter model and offers up to 100x speedup compared to standard Transformers deployments. Engineers can control generation randomness via temperature and stop sequences using simple API calls.written, reviewed, cited

GPT-Neo checkpoints are trained exclusively on the Pile dataset, so performance degrades on text distributions different from its training corpus.
The Accelerated Inference API supports inference on over 10,000 publicly available models or private models via simple API calls.
Regulations like GDPR require that users receive an explanation for automatic decisions made about them by the system.

Analysis7

GPT Neo: open-source GPT model, with pretrained 1.3B & 2.7B weight models

github.com2021-03-21authorityHN front pageabout the gpt neo familyHN 561 pts · 127 comments2,997 words

The GPT-Neo repository offers an archival implementation of model parallel GPT-3-style models using the mesh-tensorflow library, superseded by GPT-NeoX for GPU efficiency. As of August 2021, the code is no longer maintained. Two pretrained models trained on The Pile are available in 1.3B and 2.7B sizes.written, reviewed, cited

Training and inference is officially supported on TPU and should work on GPU as well.
Google colab provides tpu-v8s for free, which should be enough to finetune our models up to GPT3XL (1.5B parameter) sizes.
Some users have reported having difficulty getting MTF to recognize their GPUs.

Using GPT-J and GPT-Neo, GPT-3 open-source alternatives, with few-shot learning

nlpcloud.io2021-11-10about the gpt neo familyHN 24 pts · 1 comments

Hugging Face Transformers – Big Model Inference, Bloom, GPT Neo-X, LongT5 etc.

github.com2022-06-16authorityabout the gpt neo familyHN 2 pts · 0 comments3,146 words

GPT-Neo vs. GPT-3: Are Commercialized NLP Models That Much Better?

medium.com2021-04-09about the gpt neo familyHN 2 pts · 2 comments

GPT-3's free alternative GPT-Neo is something to be excited about

venturebeat.com2021-05-15about the gpt neo familyHN 5 pts · 0 comments

GPT-Neo on a M1 (macOS 12)

lazycoder.ro2022-02-03about the gpt neo familyHN 5 pts · 0 comments557 words

GPT-3 open-source alternatives: GPT-Neo and GPT-J

nlpcloud.io2021-12-14about the gpt neo familyHN 4 pts · 0 comments

Community1

H3 – Outperforming GPT-Neo-2.7B with only 2 attention layers

twitter.com2023-01-23HN 10 pts · 0 comments

All coverage tracked for this model (15) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
How to monitor the performance of a fine-tuned GPT-Neo Model in Production	blog.cerebrium.ai	2023-02-28	2026-10-07	hn
H3 – Outperforming GPT-Neo-2.7B with only 2 attention layers	twitter.com	2023-01-23	2026-10-07	hn
Hugging Face Transformers – Big Model Inference, Bloom, GPT Neo-X, LongT5 etc.	github.com	2022-06-16	2026-10-07	hn
GPT-Neo on a M1 (macOS 12)	lazycoder.ro	2022-02-03	2026-10-07	hn
GPT-3 open-source alternatives: GPT-Neo and GPT-J	nlpcloud.io	2021-12-14	2026-10-07	hn
Using GPT-J and GPT-Neo, GPT-3 open-source alternatives, with few-shot learning	nlpcloud.io	2021-11-10	2026-10-07	hn
Few-shot learning NLP with GPT-J and GPT-Neo	news.ycombinator.com	2021-07-22	2026-10-07	hn
Pretrain GPT-Neo for Open Source GitHub Copilot Model	discuss.huggingface.co	2021-07-03	2026-10-07	hn
Few-Shot Learning in Practice: GPT-Neo & 'HuggingFace' Accelerated Inference API	huggingface.co	2021-06-04	2026-10-07	hn
NovelAI, a new AI storyteller being built on GPT-Neo	novelai.medium.com	2021-05-31	2026-10-07	hn
GPT-3's free alternative GPT-Neo is something to be excited about	venturebeat.com	2021-05-15	2026-10-07	hn
GPT-Neo vs. GPT-3: Are Commercialized NLP Models That Much Better?	medium.com	2021-04-09	2026-10-07	hn
How to run inference with GPT-Neo for free	twitter.com	2021-04-05	2026-10-07	hn
GPT Neo: open-source GPT model, with pretrained 1.3B & 2.7B weight models	github.com	2021-03-21	2026-10-07	hn
GPT-Neo – Building a GPT-3-sized model, open source and free	eleuther.ai official	2021-01-18	2026-10-07	hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

The Pile: An 800GB Dataset of Diverse Text for Language Modeling cited in card

Leo Gao, Stella Biderman, Sid Black +9 · 2020-12-31 · PDF · HF paper page (10 upvotes)

Similar models

3 models

Similar models

phi-2 · 2.78B
same size band (M), same task (text-generation), more providers, permissive license
gemma-2-2b-it · 2.61B
same size band (M), same task (text-generation), more providers
gemma-2-2b · 2.61B
same size band (M), same task (text-generation), more providers

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:38 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:09 UTC · Last verified 2026-10-25

Follow gpt-neo-2.7B

Watch this model

/feeds/models/EleutherAI--gpt-neo-2.7B.xml(RSS 2.0) · every change to gpt-neo-2.7B
/feeds/families/EleutherAI--gpt-neo.xml(RSS 2.0) · every change across the gpt-neo family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.