Note #1 — Attention Is All You Need (Google, 2017). The paper that introduced the transformer.

Note #2 — PyTorch (Meta, 2019). The framework open AI runs on.

google-bert/bert-base-uncased

Fill-Mask • 0.1B • Updated Feb 19, 2024 • 65M • • 2.7k

Note #3 — BERT (Google, 2018). The model that kicked off the open model era and the Hub itself.

openai-community/gpt2

Text Generation • 0.1B • Updated Feb 19, 2024 • 13.3M • 3.33k

Note #4 — GPT-2 (OpenAI, 2019). The original open LLM.

Note #5 — word2vec: Efficient Estimation of Word Representations (Google, 2013).

Note #6 — HuggingFace's Transformers: State-of-the-art NLP (2019). The library open AI runs on — with its siblings Datasets, Hub, and TGI.

Note #7 — ULMFiT (Howard & Ruder, fast.ai, 2018). Universal Language Model Fine-tuning — transfer learning for NLP before BERT.

meta-llama/Llama-2-7b

Text Generation • Updated Apr 17, 2024 • 116 • 4.51k

Note #8 — Llama 2 (Meta, 2023). Open weights went commercial and the ecosystem exploded.

openai/whisper-large-v3

Automatic Speech Recognition • 2B • Updated Aug 12, 2024 • 5.87M • • 5.91k

Note #9 — Whisper (OpenAI). Open speech recognition for the whole world.

openai/gpt-oss-120b

Text Generation • 120B • Updated Aug 26, 2025 • 4.25M • • 4.95k

Note #10 — gpt-oss-120b (OpenAI, 2025). OpenAI's return to open weights, Apache 2.0.

Note #11 — ImageNet (Fei-Fei Li, Princeton/Stanford). The dataset that launched deep learning.

Note #12 — TensorFlow (Google, 2016). Large-scale ML, open sourced.

openai/clip-vit-large-patch14

Zero-Shot Image Classification • 0.4B • Updated Sep 15, 2023 • 12.6M • 2.05k

Note #13 — CLIP (OpenAI, 2021). Connected vision and language; powers image generation everywhere.

Note #14 — GPT-3: Language Models are Few-Shot Learners (OpenAI, 2020). In-context learning arrives.

Note #15 — ELMo: Deep Contextualized Word Representations (AI2/UW, 2018). Contextual embeddings arrive.

stanfordnlp/glove

Updated Aug 14, 2021 • 9

Note #16 — GloVe (Stanford, 2014). Global word vectors; with word2vec, the embeddings that started it all.

meta-llama/Meta-Llama-3-8B

Text Generation • 8B • Updated Sep 27, 2024 • 1.31M • • 6.59k

Note #17 — Llama 3 (Meta, 2024). The most downloaded LLM family in Hub history.

Note #18 — SQuAD (Stanford, 2016). The QA benchmark that defined an era of NLP.

Note #19 — LLaMA: Open and Efficient Foundation Language Models (Meta, 2023).

distilbert/distilbert-base-uncased

Fill-Mask • 67M • Updated May 6, 2024 • 8.54M • • 908

Note #20 — DistilBERT (Hugging Face). Distillation for the masses; billions of downloads.

Note #21 — Distilling the Knowledge in a Neural Network (Hinton et al., Google, 2015).

google-t5/t5-base

Translation • 0.2B • Updated Feb 14, 2024 • 1.84M • • 778

Note #22 — T5 (Google, 2019). Text-to-text everything.

Note #23 — GLUE (NYU/UW/DeepMind-adjacent, led from NYC). The benchmark suite of the BERT era.

FacebookAI/roberta-base

Fill-Mask • 0.1B • Updated Feb 19, 2024 • 10.3M • • 621

Note #24 — RoBERTa (Meta, 2019). BERT, robustly optimized; still a workhorse.

Note #25 — C4 (Google/AI2). The Colossal Clean Crawled Corpus behind T5 and countless LLMs.

Note #26 — Scaling Laws for Neural Language Models (Kaplan et al., OpenAI, 2020).

Note #27 — Adam: A Method for Stochastic Optimization (2015). The optimizer that trains almost everything on this list.

Note #28 — Batch Normalization (Google, 2015). The trick that made very deep networks trainable.

Note #29 — Generative Adversarial Networks (Goodfellow et al., 2014). The paper that launched generative modeling.

Note #30 — StyleGAN (NVIDIA, 2018). The GAN era's crown jewel; 'this person does not exist.'

Note #31 — Denoising Diffusion Probabilistic Models (UC Berkeley, 2020). DDPM.

Note #32 — Scalable Diffusion Models with Transformers — DiT (Berkeley/NYU, 2022). The backbone of modern image and video generation.

Note #33 — Deep Residual Learning for Image Recognition (Microsoft, 2015).

Note #34 — EfficientNet (Google, 2019). Compound scaling for vision.

Note #35 — The Open LLM Leaderboard. Set the pace of the open model race.

Note #36 — The Pile (EleutherAI, 2020). 825GB that trained a generation of open LLMs.

Note #37 — InstructGPT (OpenAI, 2022). RLHF as we know it.

Note #38 — GPT-4 Technical Report (OpenAI, 2023).

Note #39 — MMLU (Hendrycks et al., Berkeley). The default knowledge benchmark for years.

bigcode/starcoder

Text Generation • 16B • Updated Oct 8, 2024 • 22.8k • 2.97k

Note #40 — StarCoder (BigCode: Hugging Face + ServiceNow). Open code generation at scale.

EleutherAI/gpt-j-6b

Text Generation • Updated Jun 21, 2023 • 257k • 1.53k

Note #41 — GPT-J (EleutherAI, 2021). The grassroots open GPT-3 rival.

meta-llama/Llama-3.1-405B

Text Generation • 406B • Updated Sep 25, 2024 • 212k • • 978

Note #42 — Llama 3.1 405B (Meta, 2024). The first open frontier-scale model.

bigscience/bloom

Text Generation • 176B • Updated Jul 28, 2023 • 5.21k • 5.02k

Note #43 — BLOOM (BigScience, 2022). 176B params, 1000+ researchers, led by Hugging Face. The first open LLM at GPT-3 scale.

Note #44 — Evaluate & Evaluation on the Hub (Hugging Face, 2022). Open evaluation infrastructure for the Hub.

Note #45 — Model Cards for Model Reporting (Mitchell et al., Google, 2018). The documentation standard on every Hub repo.

Note #46 — Gradio (2019). The interface layer of open ML — every Space runs on it.

Note #47 — Chatbot Arena / LMArena (LMSYS, UC Berkeley). Community model rankings via millions of human votes.

google/gemma-7b

Text Generation • 9B • Updated Jun 27, 2024 • 29.6k • • 3.37k

Note #48 — Gemma (Google, 2024). Google returns to open weights.

Note #49 — Learning Transferable Visual Models From Natural Language Supervision — CLIP (OpenAI).

Note #50 — Datasets: A Community Library for NLP (Hugging Face, 2021). The library serving every dataset on this list.

xai-org/grok-1

Text Generation • Updated Mar 28, 2024 • 183 • 2.42k

Note #51 — Grok-1 (xAI, 2024). 314B parameters, the largest open release of its time.

Note #52 — FineWeb (Hugging Face, NYC-HQ). The modern open pretraining corpus.

Note #53 — The Stack (BigCode). 6TB of permissively licensed source code.

Note #54 — Chain-of-Thought Prompting (Google, 2022).

nvidia/Llama-3.1-Nemotron-70B-Instruct-HF

Text Generation • 71B • Updated Apr 13, 2025 • 24.5k • • 2.07k

Note #55 — Nemotron 70B (NVIDIA, 2024). Open post-training done right.

Note #56 — LoRA: Low-Rank Adaptation (Microsoft, 2021). Made fine-tuning affordable.

Note #57 — HH-RLHF (Anthropic). The canonical open human preference dataset.

facebook/bart-large

Feature Extraction • Updated Jun 3, 2022 • 131k • • 201

Note #58 — BART (Meta, 2019). Denoising seq2seq; summarization staple.

Note #59 — GSM8K (OpenAI). Grade school math that humbled LLMs for years.

Note #60 — FlashAttention (Stanford, 2022). Made long context practical.

EleutherAI/gpt-neox-20b

Text Generation • 21B • Updated Jan 31, 2024 • 420k • 584

Note #61 — GPT-NeoX-20B (EleutherAI, 2022). Open scale before it was cool.

Note #62 — RedPajama 1T (Together AI). Open reproduction of the LLaMA training data.

Note #63 — Direct Preference Optimization (Stanford, 2023). RLHF without the RL.

microsoft/phi-2

Text Generation • 3B • Updated Dec 8, 2025 • 838k • 3.48k

Note #64 — Phi-2 (Microsoft, 2023). Small models, textbook data, big results.

openai/gpt-oss-20b

Text Generation • 22B • Updated Aug 26, 2025 • 6.96M • • 4.77k

Note #65 — gpt-oss-20b (OpenAI, 2025). Frontier reasoning on a laptop.

Note #66 — HumanEval (OpenAI). THE code generation benchmark.

Note #67 — Segment Anything (Meta, 2023).

facebook/sam-vit-huge

Mask Generation • 0.6B • Updated Jan 11, 2024 • 453k • 198

Note #68 — SAM (Meta, 2023). Segment anything, literally.

google/vit-base-patch16-224

Image Classification • 86.6M • Updated Sep 5, 2023 • 5.51M • • 981

Note #69 — ViT (Google). Transformers eat computer vision.

Note #70 — OpenWebText. The open recreation of GPT-2's training data.

lmsys/vicuna-13b-v1.5

Text Generation • Updated Mar 17, 2024 • 15.1k • • 243

Note #71 — Vicuna (LMSYS, Berkeley). The chatbot that showed open models could chat.

Note #72 — Alpaca (Stanford, 2023). 52K instructions that launched the instruction-tuning boom.

Note #73 — Constitutional AI (Anthropic, 2022).

Note #74 — In-context Learning and Induction Heads (Anthropic, 2022). Mechanistic interpretability arrives.

allenai/OLMo-2-1124-7B

7B • Updated Jan 6, 2025 • 89.3k • 68

Note #75 — OLMo 2 (AI2, Seattle). Truly open: weights, data, code, logs.

Note #76 — Dolma (AI2). 3T tokens, fully documented open pretraining data.

Salesforce/codegen-16B-mono

Text Generation • Updated Jan 31, 2025 • 348 • 126

Note #77 — CodeGen (Salesforce, 2022). The first big open code LLM, pre-StarCoder.

Note #78 — dolly-15k (Databricks). Human-written instructions, employee-sourced.

HuggingFaceH4/zephyr-7b-beta

Text Generation • 7B • Updated Oct 16, 2024 • 184k • • 1.85k

Note #79 — Zephyr (Hugging Face H4). DPO goes mainstream.

Note #80 — PagedAttention / vLLM (UC Berkeley, 2023). How open models get served.

meta-llama/Llama-3.1-8B-Instruct

Text Generation • 8B • Updated Sep 25, 2024 • 9.03M • • 6.23k

Note #81 — Llama 3.1 8B Instruct (Meta). The default open chat model of 2024-25.

togethercomputer/RedPajama-INCITE-7B-Base

Text Generation • Updated Jun 6, 2023 • 410 • 92

Note #82 — RedPajama-INCITE (Together AI, 2023). Fully open reproduction, trained on US supercomputers.

Note #83 — FineWeb-Edu (Hugging Face). Education-filtered web data, new pretraining default.

microsoft/deberta-v3-base

Fill-Mask • Updated Sep 22, 2022 • 2.57M • • 429

Note #84 — DeBERTa v3 (Microsoft). Still the encoder to beat on GLUE-style tasks.

Note #85 — OpenAI Gym (2016). The playground that standardized reinforcement learning.

facebook/wav2vec2-base-960h

Automatic Speech Recognition • 94.4M • Updated Nov 14, 2022 • 1.49M • 398

Note #86 — wav2vec 2.0 (Meta, 2020). Self-supervised speech revolution.

Note #87 — LibriSpeech (Johns Hopkins). The canonical ASR benchmark.

facebook/dinov2-base

Image Feature Extraction • 86.6M • Updated Jan 17, 2024 • 1.44M • 182

Note #88 — DINOv2 (Meta). Self-supervised vision features for everything.

Note #89 — Masked Autoencoders Are Scalable Vision Learners (Meta, 2021).

facebook/detr-resnet-50

Object Detection • 41.6M • Updated Apr 10, 2024 • 1.94M • • 959

Note #90 — DETR (Meta, 2020). Object detection, transformerized.

Note #91 — COCO (Microsoft, 2014). Common Objects in Context; vision's other bedrock.

microsoft/resnet-50

Image Classification • 25.6M • Updated Feb 13, 2024 • 589k • • 498

Note #92 — ResNet-50 (Microsoft, 2015). The backbone of a decade of computer vision.

Ultralytics/YOLOv8

Object Detection • Updated 10 days ago • 10.1k • 374

Note #93 — YOLOv8 (Ultralytics). Real-time detection, descended from Joseph Redmon's YOLO (UW).

Note #94 — You Only Look Once (Joseph Redmon, UW, 2015). Real-time object detection.

facebook/bart-large-mnli

Zero-Shot Classification • 0.4B • Updated Sep 5, 2023 • 3.29M • • 1.58k

Note #95 — BART-MNLI (Meta). The zero-shot classification workhorse of the Hub.

Note #96 — IMDB (Stanford, 2011). The hello-world of sentiment analysis.

Note #97 — MNIST (Yann LeCun, AT&T Bell Labs/NYU). Where it all started.

xlnet/xlnet-base-cased

Text Generation • Updated Jan 24, 2023 • 408k • 82

Note #98 — XLNet (Google + CMU, 2019).

Note #99 — Sequence to Sequence Learning (Sutskever et al., Google, 2014).

Note #100 — XGBoost (University of Washington, 2016). The workhorse of practical ML for a decade.

google/electra-base-discriminator

Updated Feb 29, 2024 • 39M • 133

Note #101 — ELECTRA (Google + Stanford, 2020). Pretraining efficiency pioneer.

Note #102 — WikiText (Salesforce, 2016). The classic language modeling benchmark.

Note #103 — Penn Treebank (UPenn, 1993). The dataset NLP grew up on.

Note #104 — fastText (Meta FAIR, 2016). Fast, practical open NLP for everyone.

Note #105 — SST-2 (Stanford). Sentiment treebank, GLUE staple.

openai-community/openai-gpt

Text Generation • 0.1B • Updated Feb 19, 2024 • 212k • 296

Note #106 — GPT-1 (OpenAI, 2018). The one that started the G-P-T.

google/flan-t5-xxl

11B • Updated Jul 27, 2023 • 13k • 1.29k

Note #107 — Flan-T5 (Google, 2022). Instruction tuning at scale, open.

facebook/opt-6.7b

Text Generation • Updated Jan 24, 2023 • 139k • 120

Note #108 — OPT (Meta, 2022). Open Pre-trained Transformers, with logbooks and all.

Note #109 — Common Voice (Mozilla). Crowdsourced speech in 100+ languages.

suno/bark

Text-to-Speech • Updated Oct 4, 2023 • 18.8k • 1.54k

Note #110 — Bark (Suno, 2023). Open generative text-to-audio, pre-Suno-the-app.

Note #111 — The Bark demo Space that made TTS fun.

facebook/musicgen-small

Text-to-Audio • 0.6B • Updated Nov 17, 2023 • 290k • 496

Note #112 — MusicGen (Meta, 2023). Open music generation.

openai/jukebox-1b-lyrics

Feature Extraction • Updated Nov 10, 2022 • 78 • 21

Note #113 — Jukebox (OpenAI, 2020). Music generation with singing, years early.

Note #114 — MusicGen demo (Meta). Everyone's first AI song.

Note #115 — LJ Speech. The public-domain voice behind a thousand TTS models.

microsoft/speecht5_tts

Text-to-Speech • Updated Nov 8, 2023 • 79.4k • 836

Note #116 — SpeechT5 (Microsoft). Unified speech-text pretraining.

sesame/csm-1b

Text-to-Speech • 2B • Updated Dec 1, 2025 • 290k • 2.41k

Note #117 — CSM (Sesame, 2025). Open conversational speech that sounded human.

meta-llama/Llama-4-Scout-17B-16E-Instruct

Image-Text-to-Text • 109B • Updated May 22, 2025 • 724k • • 1.32k

Note #118 — Llama 4 Scout (Meta, 2025). Natively multimodal MoE.

Note #119 — HelpSteer2 (NVIDIA). Open preference data for reward models.

microsoft/phi-4

Text Generation • 15B • Updated Nov 24, 2025 • 885k • • 2.27k

Note #120 — Phi-4 (Microsoft, 2024). Small model, big reasoning.

Note #121 — Textbooks Are All You Need (Microsoft, 2023). The phi thesis.

Note #122 — TinyStories (Microsoft). How small can a coherent LM be?

EleutherAI/pythia-12b

Text Generation • 12B • Updated Jul 9, 2024 • 141k • 145

Note #123 — Pythia (EleutherAI). The scaling suite for LLM science.

EleutherAI/gpt-neo-2.7B

Text Generation • 3B • Updated Jul 9, 2023 • 39.9k • 503

Note #124 — GPT-Neo (EleutherAI, 2021). The first open GPT-3-style replication.

cerebras/Cerebras-GPT-13B

Text Generation • Updated Nov 22, 2023 • 697 • 652

Note #125 — Cerebras-GPT (2023). Chinchilla-optimal, wafer-scale trained.

Note #126 — RedPajama V2 (Together AI). 30T tokens of open web data.

Snowflake/snowflake-arctic-instruct

Text Generation • 479B • Updated May 21, 2024 • 34.8k • 362

Note #127 — Arctic (Snowflake, 2024). Enterprise-grade open MoE.

Note #128 — Mamba: Linear-Time Sequence Modeling (CMU + Princeton, 2023).

Snowflake/snowflake-arctic-embed-m

Sentence Similarity • 0.1B • Updated Dec 13, 2024 • 468k • • 166

Note #129 — Arctic Embed (Snowflake). Open retrieval embeddings.

ibm-granite/granite-3.3-8b-instruct

Text Generation • 8B • Updated May 12, 2025 • 97.1k • 158

Note #130 — Granite (IBM). A century-old American company, Apache 2.0 models.

apple/DCLM-7B

7B • Updated Jul 26, 2024 • 137 • 833

Note #131 — DCLM (Apple, 2024). DataComp-LM, open data-centric science.

apple/OpenELM-3B

Text Generation • 3B • Updated Feb 28, 2025 • 210 • 130

Note #132 — OpenELM (Apple, 2024). Apple ships open weights.

amazon/chronos-t5-large

Time Series Forecasting • 0.7B • Updated Nov 21, 2025 • 91.9k • 178

Note #133 — Chronos (Amazon). LLMs for time series forecasting.

google/timesfm-1.0-200m

Time Series Forecasting • Updated May 17, 2024 • 145 • 828

Note #134 — TimesFM (Google). Foundation model for forecasting.

allenai/Molmo-7B-D-0924

Image-Text-to-Text • 8B • Updated Dec 15, 2025 • 30k • 567

Note #135 — Molmo (AI2, 2024). Open VLM matching closed ones, with open data.

adept/fuyu-8b

Image-Text-to-Text • 9B • Updated Nov 4, 2023 • 133k • 1.02k

Note #136 — Fuyu (Adept, 2023). Multimodal without an image encoder.

vikhyatk/moondream2

Image-Text-to-Text • 2B • Updated Sep 23, 2025 • 1.62M • 1.43k

Note #137 — Moondream (2024). The beloved tiny vision-language model.

microsoft/Florence-2-large

Image-Text-to-Text • 0.8B • Updated Aug 4, 2025 • 693k • 1.83k

Note #138 — Florence-2 (Microsoft, 2024). One small model, every vision task.

Note #139 — BLIP-2 (Salesforce, 2023). Bootstrapping VLMs from frozen parts.

Salesforce/blip2-opt-2.7b

Image-Text-to-Text • 4B • Updated Feb 3, 2025 • 658k • 446

Note #140 — BLIP-2 (Salesforce). Vision-language bridge.

Salesforce/blip-image-captioning-base

Image-to-Text • Updated Feb 3, 2025 • 1.74M • 865

Note #141 — BLIP (Salesforce). The Hub's default image captioner for years.

Note #142 — Visual Instruction Tuning — LLaVA (UW-Madison + Microsoft, 2023).

llava-hf/llava-1.5-7b-hf

Image-Text-to-Text • 7B • Updated Jun 6, 2025 • 3.14M • 367

Note #143 — LLaVA 1.5 (UW-Madison/Microsoft). Open multimodal chat for everyone.

facebook/chameleon-7b

Image-Text-to-Text • 7B • Updated Jul 23, 2024 • 228k • 202

Note #144 — Chameleon (Meta, 2024). Early-fusion multimodal.

google/paligemma-3b-pt-224

Image-Text-to-Text • 3B • Updated Sep 21, 2024 • 276k • 495

Note #145 — PaliGemma (Google, 2024). Open VLM building block.

google/siglip-so400m-patch14-384

Zero-Shot Image Classification • 0.9B • Updated Sep 26, 2024 • 1.77M • 680

Note #146 — SigLIP (Google). The CLIP successor inside most modern VLMs.

Note #147 — MMMU (Ohio State-led). The multimodal exam.

Note #148 — LMSYS-Chat-1M (Berkeley). One million real conversations.

Note #149 — Judging LLM-as-a-Judge with MT-Bench (LMSYS, Berkeley, 2023).

NousResearch/Hermes-3-Llama-3.1-8B

Text Generation • 8B • Updated Sep 8, 2024 • 359k • • 465

Note #150 — Hermes 3 (Nous Research). The American fine-tuning scene at its best.

Note #151 — OpenHermes 2.5 (Teknium). The community SFT dataset.

Note #152 — OpenOrca. Open reproduction of Microsoft's Orca data.

berkeley-nest/Starling-LM-7B-alpha

Text Generation • 7B • Updated Mar 20, 2024 • 1.83k • • 560

Note #153 — Starling (Berkeley). Open RLAIF.

allenai/Llama-3.1-Tulu-3-8B

Text Generation • 8B • Updated Feb 13, 2025 • 3.72k • • 179

Note #154 — Tulu 3 (AI2). Fully open post-training recipe.

Note #155 — Tulu 3 SFT mixture (AI2). The open post-training data stack.

Note #156 — QLoRA (University of Washington, 2023). Fine-tune 65B on one GPU.

Note #157 — AWQ (MIT, 2023). Activation-aware quantization, everywhere now.

bigcode/starcoder2-15b

Text Generation • 16B • Updated Jun 5, 2024 • 10.9k • 674

Note #158 — StarCoder2 (BigCode, 2024).

codellama/CodeLlama-7b-hf

Text Generation • 7B • Updated Apr 12, 2024 • 242k • 377

Note #159 — Code Llama (Meta, 2023). Open code models at every size.

Note #160 — Evaluating LLMs Trained on Code — Codex/HumanEval (OpenAI, 2021).

Note #161 — MBPP (Google). Mostly Basic Python Problems.

Note #162 — SWE-bench (Princeton). Can models fix real GitHub issues?

Note #163 — SWE-bench paper (Princeton, 2023). Defined agentic coding evaluation.

Note #164 — Summarize-from-feedback (OpenAI). Foundational RLHF data.

Note #165 — Learning to Summarize from Human Feedback (OpenAI, 2020).

Note #166 — Deep RL from Human Preferences (Christiano et al., OpenAI, 2017). RLHF's origin.

Note #167 — Proximal Policy Optimization (OpenAI, 2017). The algorithm behind RLHF.

Note #168 — WebGPT comparisons (OpenAI). Early human feedback data.

Note #169 — Let's Verify Step by Step (OpenAI, 2023). Process reward models.

Note #170 — MATH-500. The competition math eval, from OpenAI's split of Hendrycks MATH.

Note #171 — PaLM: Scaling Language Modeling with Pathways (Google, 2022).

Note #172 — Switch Transformers (Google, 2021). MoE at trillion scale.

Note #173 — Outrageously Large Neural Networks (Shazeer et al., Google, 2017). The mixture-of-experts layer behind today's MoEs.

Note #174 — Emergent Abilities of Large Language Models (Google, 2022).

Note #175 — Training a Helpful and Harmless Assistant with RLHF (Anthropic, 2022).

Note #176 — Retrieval-Augmented Generation (Meta, 2020). RAG.

Note #177 — ReAct: Synergizing Reasoning and Acting (Princeton + Google, 2022).

Note #178 — Toolformer (Meta, 2023). LLMs learn to use tools.

Note #179 — Generative Agents: Interactive Simulacra (Stanford, 2023). The Smallville paper.

Note #180 — WildChat (AI2). Real user-LLM conversations, opened.

Note #181 — FLORES (Meta). The multilingual translation benchmark.

facebook/nllb-200-distilled-600M

Translation • Updated Feb 14, 2024 • 1.92M • 929

Note #182 — NLLB (Meta, 2022). No Language Left Behind: 200 languages.

facebook/seamless-m4t-v2-large

Automatic Speech Recognition • 2B • Updated Jan 4, 2024 • 161k • 990

Note #183 — SeamlessM4T (Meta, 2023). Universal speech translation.

facebook/mms-1b-all

Automatic Speech Recognition • 1.0B • Updated Jun 15, 2023 • 300k • 202

Note #184 — MMS (Meta, 2023). Speech tech for 1,000+ languages.

distil-whisper/distil-large-v3

Automatic Speech Recognition • 0.8B • Updated Apr 21 • 776k • 376

Note #185 — Distil-Whisper (Hugging Face). Whisper, 6x faster.

nvidia/parakeet-tdt-0.6b-v2

Automatic Speech Recognition • Updated 7 days ago • 363k • 1.51k

Note #186 — Parakeet (NVIDIA, 2025). Top of the open ASR leaderboard.

nvidia/canary-1b

Automatic Speech Recognition • Updated Dec 3, 2025 • 2.22k • 457

Note #187 — Canary (NVIDIA). Multilingual ASR and translation.

Note #188 — The People's Speech (MLCommons). 30K+ hours of open ASR data.

Note #189 — Open ASR Leaderboard. Keeping speech recognition honest.

Note #190 — The Whisper demo Space. Transcription for everyone.

nvidia/NV-Embed-v2

Feature Extraction • 8B • Updated Jul 21, 2025 • 25.3k • 513

Note #191 — NV-Embed (NVIDIA). Top open embedding model of its day.

nomic-ai/nomic-embed-text-v1.5

Sentence Similarity • 0.1B • Updated Apr 7 • 16.1M • 860

Note #192 — Nomic Embed (NYC). Open weights, open data embeddings.

Note #193 — MTEB Leaderboard. Where embedding models compete.

answerdotai/ModernBERT-base

Fill-Mask • 0.1B • Updated Jan 15, 2025 • 10M • 1.07k

Note #194 — ModernBERT (Answer.AI, 2024). BERT, rebuilt for the 2020s.

allenai/longformer-base-4096

Updated Apr 5, 2023 • 1.48M • 229

Note #195 — Longformer (AI2, 2020). Long documents before long context.

microsoft/biogpt

Text Generation • Updated Feb 3, 2023 • 123k • 307

Note #196 — BioGPT (Microsoft). Biomedical text generation.

facebook/esm2_t33_650M_UR50D

Fill-Mask • 0.7B • Updated Mar 21, 2023 • 1.2M • • 82

Note #197 — ESM-2 (Meta). Protein language models, open sourced — Meta's open answer to AlphaFold.

facebook/esmfold_v1

Updated Mar 22, 2023 • 1.58M • 50

Note #198 — ESMFold (Meta, 2022). Open protein structure prediction — Meta's open answer to AlphaFold.

Note #199 — peS2o (AI2). Open academic papers corpus.

facebook/galactica-6.7b

Text Generation • Updated Jan 24, 2023 • 694 • 104

Note #200 — Galactica (Meta, 2022). Ahead of its time, in every sense.

Note #201 — Objaverse (AI2). 800K+ annotated 3D objects.

openai/shap-e

Text-to-3D • Updated Dec 11, 2023 • 3.26k • 277

Note #202 — Shap-E (OpenAI, 2023). Open text-to-3D.

Note #203 — NeRF: Neural Radiance Fields (Berkeley + Google, 2020).

genmo/mochi-1-preview

Text-to-Video • Updated Sep 4, 2025 • 2.86k • • 1.33k

Note #204 — Mochi (Genmo, 2024). Open video generation from the Bay.

playgroundai/playground-v2.5-1024px-aesthetic

Text-to-Image • Updated Mar 15, 2024 • 306k • • 767

Note #205 — Playground v2.5. American open text-to-image.

fal/AuraFlow

Text-to-Image • Updated Jul 18, 2024 • 305 • • 654

Note #206 — AuraFlow (fal). Open flow-based image generation.

Note #207 — DALL-E mini (Boris Dayma, Texas). The Space that broke the internet in 2022.

Note #208 — DALL-E 2: Hierarchical Text-Conditional Image Generation (OpenAI, 2022).

Note #209 — Diffusion Models Beat GANs on Image Synthesis (OpenAI, 2021).

Note #210 — Conceptual Captions (Google). Image-text pretraining pioneer.

Note #211 — DiffusionDB (Georgia Tech). 14M images from Stable Diffusion users.

Note #212 — QR Code AI Art. Functional QR codes as art; a 2023 moment.

Note #213 — AI Comic Factory. One of the most-liked Spaces ever.

Note #214 — HuggingChat. Open source ChatGPT alternative.

HuggingFaceTB/SmolLM3-3B

Text Generation • 3B • Updated Sep 10, 2025 • 698k • 981

Note #215 — SmolLM3 (Hugging Face, 2025). Fully open small model, trained in the open.

Note #216 — Cosmopedia (Hugging Face). Largest open synthetic pretraining dataset.

LiquidAI/LFM2-1.2B

Text Generation • 1B • Updated Feb 12 • 144k • 355

Note #217 — LFM2 (Liquid AI, MIT spinoff, 2025). Post-transformer architectures, open.

PrimeIntellect/INTELLECT-1

Text Generation • 10B • Updated Nov 29, 2024 • 119 • 67

Note #218 — INTELLECT-1 (Prime Intellect, 2024). First decentralized-trained 10B model.

microsoft/bitnet-b1.58-2B-4T

Text Generation • 0.8B • Updated Dec 17, 2025 • 8.6k • 1.47k

Note #219 — BitNet (Microsoft, 2025). 1.58-bit LLMs, for real.

nvidia/Llama-3_1-Nemotron-Ultra-253B-v1

Text Generation • 253B • Updated Oct 15, 2025 • 1.03k • 352

Note #220 — Nemotron Ultra 253B (NVIDIA, 2025). Open reasoning at scale.

nvidia/GR00T-N1-2B

Robotics • 2B • Updated Sep 2, 2025 • 202 • 356

Note #221 — GR00T N1 (NVIDIA, 2025). Open foundation model for humanoid robots.

lerobot/smolvla_base

Robotics • 0.5B • Updated Jan 22 • 49.4k • 396

Note #222 — SmolVLA (Hugging Face LeRobot, 2025). Open vision-language-action model for robotics.

openvla/openvla-7b

Robotics • 8B • Updated Feb 17 • 1.28M • 237

Note #223 — OpenVLA (Stanford, 2024). Open vision-language-action for robots.

meta-llama/Llama-3.3-70B-Instruct

Text Generation • 71B • Updated Dec 21, 2024 • 763k • • 2.88k

Note #224 — Llama 3.3 70B (Meta). 405B-class quality, 70B price.

meta-llama/Llama-Guard-3-8B

Text Generation • 8B • Updated Oct 11, 2024 • 243k • • 308

Note #225 — Llama Guard (Meta). Open safety classification.

google/gemma-3-27b-it

Image-Text-to-Text • 27B • Updated Mar 21, 2025 • 870k • • 1.99k

Note #226 — Gemma 3 (Google, 2025). Multimodal, 140+ languages, single GPU.

Note #227 — Gemma: Open Models Based on Gemini Research (Google, 2024).

Note #228 — CNN/DailyMail (Stanford/Google). The summarization benchmark.

Note #229 — MS MARCO (Microsoft). The dataset behind neural search.

Note #230 — Natural Questions (Google). Real queries, real answers.

Note #231 — HellaSwag (UW/AI2). Commonsense that fooled models for years.

Note #232 — WinoGrande (AI2). Winograd schemas at scale.

Note #233 — ARC (AI2). The AI2 Reasoning Challenge.

Note #234 — SuperGLUE (NYU/UW). GLUE's harder sibling.

Note #235 — TruthfulQA (OpenAI-affiliated, 2021). Measuring imitative falsehoods.

Note #236 — IFEval (Google). Instruction-following, verifiable.

Note #237 — GPQA (NYU). Google-proof graduate-level science questions.

Note #238 — Humanity's Last Exam (CAIS + Scale AI, 2025). The frontier benchmark.

Note #239 — GAIA (Meta + Hugging Face). The agent benchmark.

Note #240 — RewardBench (AI2). Evaluating reward models.

Note #241 — BigCode Models Leaderboard. Open code model rankings.

Note #242 — Megatron-LM (NVIDIA, 2019). Model parallelism playbook.

Note #243 — ZeRO (Microsoft, 2019). DeepSpeed's memory magic.

Note #244 — The FineWeb Datasets (Hugging Face, 2024).

Note #245 — OpenAI o1 System Card (2024). Reasoning models arrive.

allenai/OLMo-7B

Text Generation • 7B • Updated Oct 9, 2025 • 55.3k • 653

Note #246 — OLMo 1 (AI2, 2024). The first fully open LLM pipeline.

allenai/olmOCR-7B-0225-preview

Image-Text-to-Text • 8B • Updated Aug 19, 2025 • 2.39k • 708

Note #247 — olmOCR (AI2, 2025). Open document understanding hit.

microsoft/layoutlmv3-base

0.1B • Updated Apr 10, 2024 • 856k • 501

Note #248 — LayoutLMv3 (Microsoft). Document AI foundation.

Note #249 — DeepSite. Vibe coding in a Space; 2025 phenomenon.

Note #250 — gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, 2025).