DeepSeek-OCR and GPT-2 share Hugging Face's top-likes board

The most-liked models page on Hugging Face is a strange snapshot of what the open model community has actually decided to bookmark. Stability AI's stable-video-diffusion-img2vid-xt sits at the top with 3.33k likes, OpenAI's GPT-2 is essentially tied a step behind at 3.32k, and DeepSeek-OCR — listed as updated November 4, 2025 — is already at 3.29k likes and 2.21M downloads after only weeks on the platform.

What stands out is how mixed the era is. CompVis's stable-diffusion-v-1-4-original, last updated in November 2022, is still in the top thirty by likes. So is prompthero's openjourney from May 2023, and Google's BERT-base-uncased, which has 61.2 million downloads to its name on a page that lists it as last touched in February 2024. The leaderboard reads less like a recency feed and more like an accumulated hall of fame.

The new entrants tell their own story. DeepSeek shows up three times in the visible list (OCR, V3-0324, and R1-0528), Mistral twice (7B-Instruct v0.2 and v0.3), and Meta twice (Llama 3.3 70B and Llama 3.2 1B). BAAI's bge-m3 embedding model is at 31.1 million downloads, OpenAI's whisper-large-v3-turbo has logged 6.98 million, and pyannote's speaker-diarization-3.1 has 8.07 million. The pattern is that the most-loved checkpoints are either workhorses you bolt into a pipeline (embeddings, ASR, diarization, fill-mask, OCR) or the open frontier text-generation models people actually try to run themselves.

The honest caveat is that likes on Hugging Face is a vanity metric. It does not separate a research pull from a production deployment, and it heavily favors anything that has been on the site for years. A model that landed last month is competing against ones that have had three years to collect bookmarks. The page also gives no view of which checkpoints are gaining attention fastest, only cumulative totals. Take the rankings as a directory of attention, not a benchmark of quality.

If you are building, the useful read is which categories are over-indexed. Open text generation, retrieval embeddings, speech recognition, OCR, and image and video generation are where the community is concentrating its likes. The orgs benefiting most from that attention right now look like DeepSeek, Mistral, Meta, BAAI and Stability, and the long tail of older Stability and CompVis releases is not going away soon.