Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › gpt2 family

gpt2-xl

Hugging Face ↗ Follow changes

openai-community MIT out 5 years ago

1.61B parameters, text in, text out, 1K tokens of context.

Verified 2026-10-25 Rank #573 in LLMs · score 74.4 method v1 Sources 3 Downloads 30d 73.8k · #350 of 619 tracked

Can I run it?1.0GB at INT4Estimatefits a 16 GB cardmethodCommunity: Show HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colab →
Can I use it?MITPermissivelicence file2026-10-24
Do people use it?73.8kdownloads, 30 days#350 of 619 we track · below the median+1 likes, downloads +6% this week393 likesHugging Face2026-10-24

Worth reading

Picked by rule from 9 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 7 months ago

OpenAI says its new model GPT-2 is too dangerous to release (2019)

On Feb.

  • OpenAI withheld the full algorithm and training codes despite releasing a much smaller version of the model.
  • The model distinguishes between multiple definitions of a single word based on context clues and has deeper knowledge of obscure usages.
slate.com 2026-04-08 395 HN pts · 120 comments 1.7k words about the gpt2 family family

9 best of 9 tracked

The 2 picks above, plus these 7. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeShow HN: GPT-2 implemented using graphics shadersgithub.com · 2025-05-02 · 228 HN pts
  2. ArticleTry to guess if code is real or GPT2-generateddoesnotexist.codes · 2021-02-23 · 201 HN pts
  3. ArticleTalking to myself: how I trained GPT2-1.5b for rubber ducking using my chat datasvilentodorov.xyz · 2020-01-23
  4. ArticleShow HN: Fully client-side GPT2 prediction visualizerperplexity.vercel.app · 2023-09-05 · 153 HN pts
  5. ArticleThe Illustrated GPT-2: Visualizing Transformer Language Models (2019)jalammar.github.io · 2023-12-19 · 213 HN pts
  6. CodeShow HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colabgithub.com · 2023-02-13 · 5 HN pts
  7. ArticleSteering GPT-2-XL by adding an activation vectorlesswrong.com · 2023-05-14 · 5 HN pts

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: 20 quantized builds and 100 Spaces.

  • As a service0no API provider listed
  • On own hardware20quantized builds
  • Built on23368 fine-tunes · 165 adapters · 0 merges
  • In apps100+Hugging Face Spaces

Top Spaces running it: IndexTeam/IndexTTS-2-Demo ♥ 849 · pliny-the-prompter/obliteratus ♥ 621 · mrfakename/MetaVoice-1B-v0.1 ♥ 144 · nateraw/lavila ♥ 66

Will it fit?

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

Fits a 16 GB card at INT4

About 1.0 GB at INT4, 1.9 GB at FP8, 3.9 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.0 GB of weights (estimate; KV cache and activations extra)INT41.0 GBFP8: about 1.9 GB of weights (estimate; KV cache and activations extra)FP81.9 GBBF16: about 3.9 GB of weights (estimate; KV cache and activations extra)BF163.9 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.0 GB of weights (estimate; KV cache and activations extra)INT41.0 GBFP8: about 1.9 GB of weights (estimate; KV cache and activations extra)FP81.9 GBBF16: about 3.9 GB of weights (estimate; KV cache and activations extra)BF163.9 GB

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

The story so far

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

8 dated articles and 21 repo commits so far

Teal is the weights (2022-03-02), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.

DecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2024-02-19: 1 repo commit2023-10-23: 1 repo commit2023-01-24: 1 repo commit2022-08-23: 1 repo commit2022-03-08: 1 repo commit2021-09-08: 1 repo commit2021-05-21: 1 repo commit2020-12-11: 1 repo commit2020-12-09: 1 repo commit2020-10-12: 1 repo commit2020-05-11: 1 repo commit2020-05-02: 1 repo commit2020-04-24: 1 repo commit2020-03-13: 1 repo commit2020-01-31: 1 repo commit2019-12-20: 1 repo commit2019-11-05: 5 repo commitsOpenAI says its new model GPT-2 is too dangerous to release (2019) · 2026-04-08Bumblebee: GPT2, Stable Diffusion, and More in Elixir · 2022-12-08Show HN: GPT-2 implemented using graphics shaders · 2025-05-02Try to guess if code is real or GPT2-generated · 2021-02-23Show HN: Fully client-side GPT2 prediction visualizer · 2023-09-05The Illustrated GPT-2: Visualizing Transformer Language Models (2019) · 2023-12-19Show HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colab · 2023-02-13Steering GPT-2-XL by adding an activation vector · 2023-05-14Weights 2022-03-02 · 2022-03-02Weights 2022-03-02now
DecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2024-02-19: 1 repo commit2023-10-23: 1 repo commit2023-01-24: 1 repo commit2022-08-23: 1 repo commit2022-03-08: 1 repo commit2021-09-08: 1 repo commit2021-05-21: 1 repo commit2020-12-11: 1 repo commit2020-12-09: 1 repo commit2020-10-12: 1 repo commit2020-05-11: 1 repo commit2020-05-02: 1 repo commit2020-04-24: 1 repo commit2020-03-13: 1 repo commit2020-01-31: 1 repo commit2019-12-20: 1 repo commit2019-11-05: 5 repo commitsOpenAI says its new model GPT-2 is too dangerous to release (2019) · 2026-04-08Bumblebee: GPT2, Stable Diffusion, and More in Elixir · 2022-12-08Show HN: GPT-2 implemented using graphics shaders · 2025-05-02Try to guess if code is real or GPT2-generated · 2021-02-23Show HN: Fully client-side GPT2 prediction visualizer · 2023-09-05The Illustrated GPT-2: Visualizing Transformer Language Models (2019) · 2023-12-19Show HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colab · 2023-02-13Steering GPT-2-XL by adding an activation vector · 2023-05-14Weights 2022-03-02 · 2022-03-02Weights 2022-03-02now

What people built on it

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

253 derivatives on the Hub, most of them adapters

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

20 quantized (Hugging Face model tree)68 fine-tunes (Hugging Face model tree)165 adapters (Hugging Face model tree)20 quantized68 fine-tunes165 adapters
20 quantized (Hugging Face model tree)68 fine-tunes (Hugging Face model tree)165 adapters (Hugging Face model tree)20 quantized68 fine-tunes165 adapters
GGUF DevQuasar-10/openai-community.gpt2-xl-GGUF for llama.cpp, Ollama, LM Studio 12.2k downloads MLX MCES10/gpt2-xl-mlx-fp16 for Apple silicon 918 downloads

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.
BenchmarkValueSetting / notesReported byDate
author-reported 1BW42.16 PPLzero-shotmodel author2026-10-07
author-reported CBT-CN93.3 ACCzero-shotmodel author2026-10-07
author-reported CBT-NE89.05 ACCzero-shotmodel author2026-10-07
author-reported LAMBADA8.63 PPLzero-shotmodel author2026-10-07
author-reported LAMBADA63.24 ACCzero-shotmodel author2026-10-07
author-reported PTB35.76 PPLzero-shotmodel author2026-10-07
author-reported WikiText10317.48 PPLzero-shotmodel author2026-10-07
author-reported WikiText218.34 PPLzero-shotmodel author2026-10-07
author-reported enwiki80.93 BPBzero-shotmodel author2026-10-07
author-reported text80.98 BPCzero-shotmodel author2026-10-07

What the card says

2 cited sentences

Summary generated, cited

GPT-2 XL is a 1.61B parameter transformer-based language model designed for text generation.[1][2] The model is pretrained on English using a causal language modeling objective.[3][4]

  1. parameters human
  2. “Transformer-based language model”
  3. languages
  4. “pretrained model on English language using a causal langu…”

Generated 2026-10-07 15:13 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

1 use case

What it is good at 1 use case

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.

Facts

20 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesGPT2LMHeadModel
Model typegpt2
Params by dtypeF32: 1,607,942,848
Size bandS
MoEno
Context length1,024 (config.json max_position_embeddings)
Languagesen
PapersarXiv:1910.09700
LicenseMIT (mit)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · GPT2LMHeadModel · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · GPT2LMHeadModel · 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 gpt2-xl: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)3.9 GB · Edge (<4 GB)
FP8 (1 byte/param)1.9 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.0 GB · Edge (<4 GB)

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

All variants

40 variants

Lineage

Original model Family: openai-community/gpt2 253 derivatives on the Hub

finetune: 68quantized: 20adapter: 165

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
nicholasKluge/Aira-2-1B5 finetune — 1.56B S apache-2.0 nicholasKluge 1 23.1k
DevQuasar-10/openai-community.gpt2-xl-GGUF quantized GGUF — — — DevQuasar-10 0 12.2k
MiniLLM/teacher-gpt2-1.5B finetune — — — apache-2.0 MiniLLM 3 8.4k
mradermacher/gpt2-xl-i1-GGUF quantized GGUF — — mit mradermacher 0 7.6k
isaacus/open-australian-legal-llm finetune — 1.56B S apache-2.0 isaacus 14 6.8k
mradermacher/gpt2-xl-GGUF quantized GGUF — — mit mradermacher 0 7.1k
tensorblock/gpt2-xl-GGUF quantized GGUF — — mit tensorblock 2 3.4k
Alcoft/gpt2-xl-GGUF quantized GGUF — — mit Alcoft 1 3.3k
Clover-Hill/gpt2-xl-finetuned-wikitext103 finetune — 1.56B S apache-2.0 Clover-Hill 0 2.9k
aieng-lab/gpt2-xl_incivility finetune — 1.56B S mit aieng-lab 0 2.7k
Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF quantized Q2_K — — mit Ffftdtd5dtft 0 2.3k
waveforce-ai/Aura-2-Lightning finetune — 1.56B S apache-2.0 waveforce-ai 3 1.9k
aieng-lab/gpt2-xl_story-points finetune — 1.56B S mit aieng-lab 0 1.5k
MLVXN/MicroLLM2 adapter GGUF 1.56B S apache-2.0 MLVXN 2 1.2k
lukasstraub2/gpt2-aidungeon2 finetune — 1.56B S mit lukasstraub2 1 965
MCES10/gpt2-xl-mlx-fp16 finetune mlx 1.56B S mit MCES10 0 918
warshanks/gpt2-xl-bf16 finetune -bf16 1.56B S mit warshanks 0 773
MHGanainy/gpt2-xl-lora-multi-shared-512 adapter — — — mit MHGanainy 0 772
anester/gpt2-xl-Q4_K_M-GGUF quantized GGUF — — mit anester 0 700
schaeff/gpt2-xl_LNFree800 finetune — 1.56B S other schaeff 0 660
OBLITERATUS/gpt2-xl-OBLITERATED finetune — 1.56B S — OBLITERATUS 2 504
ysn-rfd/gpt2-xl-Q8_0-GGUF quantized Q8_0 — — mit ysn-rfd 1 536
MHGanainy/gpt2-xl-lora-ecthr adapter — — — mit MHGanainy 0 436
Rubin-Wei/gpt2-large-finetuned-wikitext103 finetune — 774M S apache-2.0 Rubin-Wei 0 417
MHGanainy/gpt2-xl-lora-multi adapter — — — mit MHGanainy 0 399
marcolorenzi98/GPT2-xl-finetuned-covid_fake_news finetune — 1.56B S mit marcolorenzi98 0 347
CIRCL/vulnerability-description-generation-gpt2-xl finetune — 1.56B S mit CIRCL 3 187
mergekit-community/BetterGPT2 finetune — 1.56B S — mergekit-community 0 315
jdumanski/gpt2-xl-Q4_K_M-GGUF quantized GGUF — — mit jdumanski 0 305
MHGanainy/gpt2-xl-lora-multi-512-k5-31-im-2 adapter — — — mit MHGanainy 0 281
EntityDeletr/gpt2-xl quantized GGUF 1.61B S mit EntityDeletr 1 231
MHGanainy/gpt2-xl-lora-multi-512-4-top adapter — — — mit MHGanainy 0 275
Striker-7/results adapter — — — mit Striker-7 1 220
antoste/gpt2-xl-Q8_0-GGUF quantized Q8_0 — — mit antoste 0 267
lieren2023/gpt2-xl-Q8_0-GGUF quantized Q8_0 — — mit lieren2023 0 260
Akul/gpt2-xl-Q6_K-GGUF quantized Q6_K — — mit Akul 0 249
BananaMind/BananaMind-1.0-Instruct finetune — 1.56B S apache-2.0 BananaMind 0 239
hardrave/zombiellm finetune — 1.56B S cc-by-nc-4.0 hardrave 2 134
schaeff/gpt2-xl_vanilla800 finetune — 1.56B S other schaeff 0 232
marroyo777/gpt2-xl-Q4_K_M-GGUF quantized GGUF — — mit marroyo777 0 216

Repository changelog

21 commits

Repository changelog 21 commits

DateCommitSHA
2024-02-19 Adds the tokenizer configuration file (#9) 15ea56d
2023-10-23 Adding `safetensors` variant of this model (#5) b81cf80
2023-01-24 Adding generation config file(s) 33cdb5c
2022-08-23 Add model card (#1) 208c706
2022-03-08 upload flax 82bb810
2021-09-08 Update config.json cefd64e
2021-05-21 allow flax ba796d0
2020-12-11 Migrate model card from transformers-repo 690cfc3
2020-12-09 For clarity, delete deprecated modelcard.json e5f9794
2020-10-12 Update tokenizer.json f6d9a45
2020-05-11 Update config.json ed58fb2
2020-05-02 Update rust_model.ot 2b6e283
2020-04-24 Update config.json 972b329
2020-03-13 Update tf_model.h5 5f366f8
2020-01-31 Update config.json ff1fc14
2019-12-20 Update modelcard.json 4bcc4ab
2019-11-05 Update pytorch_model.bin 96efdb1
2019-11-05 Update merges.txt e35c404
2019-11-05 Update vocab.json 8d46d24
2019-11-05 Update config.json f5f1719
2019-11-05 initial commit 95c0a9c

Articles and coverage

9 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.

Analysis8

OpenAI says its new model GPT-2 is too dangerous to release (2019)
slate.com2026-04-08HN front pageabout the gpt2 familyHN 395 pts · 120 comments1,704 words
On Feb. 14, OpenAI announced a text-generation model trained using text from 8 million webpages that can adapt to the style and content of conditioning text. The organization decided to withhold the full algorithm and training codes due to concerns that it could be used to generate false news articles or impersonate people online. Experts note that an entity with enough capital could build a comparable generator by renting servers from Amazon Web Services.written, reviewed, cited
  • OpenAI withheld the full algorithm and training codes despite releasing a much smaller version of the model.
  • The model distinguishes between multiple definitions of a single word based on context clues and has deeper knowledge of obscure usages.
Bumblebee: GPT2, Stable Diffusion, and More in Elixir
news.livebook.dev2022-12-08HN front pageabout the gpt2 familyHN 550 pts · 72 comments
Show HN: GPT-2 implemented using graphics shaders
github.com2025-05-02authorityHN front pageabout the gpt2 familyHN 228 pts · 25 comments169 words
The project implements a full GPT-2 small (117M) forward pass in the GPU via WebGL2 shaders. BPE tokenization uses js-tiktoken in the browser without a WASM fetch.generated, cited
Try to guess if code is real or GPT2-generated
doesnotexist.codes2021-02-23HN front pageabout the gpt2 familyHN 201 pts · 103 comments
Show HN: Fully client-side GPT2 prediction visualizer
perplexity.vercel.app2023-09-05HN front pageabout the gpt2 familyHN 153 pts · 11 comments304 words
The Illustrated GPT-2: Visualizing Transformer Language Models (2019)
jalammar.github.io2023-12-19HN front pageabout the gpt2 familyHN 213 pts · 5 comments5,011 words
Show HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colab
github.com2023-02-13authorityHN 5 pts · 1 comments78 words
Steering GPT-2-XL by adding an activation vector
lesswrong.com2023-05-14HN 5 pts · 0 comments147 words
All coverage tracked for this model (9) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
OpenAI says its new model GPT-2 is too dangerous to release (2019)slate.com2026-04-082026-10-07hn
Show HN: GPT-2 implemented using graphics shadersgithub.com2025-05-022026-10-07hn
The Illustrated GPT-2: Visualizing Transformer Language Models (2019)jalammar.github.io2023-12-192026-10-07hn
Show HN: Fully client-side GPT2 prediction visualizerperplexity.vercel.app2023-09-052026-10-07hn
Steering GPT-2-XL by adding an activation vectorlesswrong.com2023-05-142026-10-07hn
Show HN: Recurser lib reduces GPT2-XL VRAM usage by 25% and runs it on Colabgithub.com2023-02-132026-10-07hn
Bumblebee: GPT2, Stable Diffusion, and More in Elixirnews.livebook.dev2022-12-082026-10-07hn
Try to guess if code is real or GPT2-generateddoesnotexist.codes2021-02-232026-10-07hn
Talking to myself: how I trained GPT2-1.5b for rubber ducking using my chat datasvilentodorov.xyz2020-01-232026-10-07hn

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).

Similar models

3 models

Similar models

  • Qwen2.5-1.5B-Instruct · 1.54B
    same size band (S), same task (text-generation), more providers, permissive license
  • Qwen2.5-1.5B · 1.54B
    same size band (S), same task (text-generation), more providers, permissive license
  • DeepSeek-R1-Distill-Qwen-1.5B · 1.78B
    same size band (S), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow gpt2-xl

Watch this model

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