OpenAI's Open Models Take Shape on Hugging Face
Welcome to the official Hugging Face organization for OpenAI's open models! 🤗 This profile documents OpenAI's presence on the platform, its open-weight releases, community activity, and the growing ecosystem of models, datasets, and spaces under its banner. As of the latest count, OpenAI maintains 39 models and 16 datasets, supported by a team of 144 people visible on the organization card. The page greets visitors with the statement: "Welcome to the official Hugging Face organization for OpenAI's open models! 🤗"
Open-Weight Language Models
OpenAI publishes several open-weight language models on Hugging Face, each with distinct parameter counts and update histories. The flagship gpt-oss-120b, with 117 billion parameters, arrived in August 2025 and has already crossed 5.17 million downloads. Its smaller sibling, gpt-oss-20b, at 21 billion parameters, reached 6.43 million downloads by the same mark. Both models carry bio descriptions: gpt-oss-120b is described as "our most advanced powerful open model designed for complex tasks, deeper context understanding, and enhanced reasoning capabilities," while gpt-oss-20b is labeled "a versatile, efficient open-source language model ideal for a wide range of applications, from conversational AI to creative content generation."
Beyond the GPT-OSS family, OpenAI's Whisper series covers automatic speech recognition across multiple sizes. The lineup includes large-v3-turbo (0.8B params, 7.15M downloads, updated October 2024), large-v3 (2B params, 4.59M downloads), large-v2 (2B params, 155k downloads), and the original large model (2B params, 39.7k downloads). The CLIP model rounds out the multimodal offerings, described as learning "visual concepts from natural language, enabling zero-shot image classification by matching text descriptions with images, improving robustness and flexibility across diverse tasks."
Safety-Tuned Derivatives
Three safety-oriented variants sit under the gpt-oss umbrella. gpt-oss-safeguard-120b (120B params, updated October 29, 2025) and gpt-oss-safeguard-20b (22B params, updated January 14) are safety reasoning models built upon the base gpt-oss models. A third variant, simply named gpt-oss-safeguard, inherits the same architectural lineage. Each carries download metrics in the tens of thousands, reflecting a more niche but strategically important subset of the portfolio.
Dataset Collection and Community Activity
Sixteen datasets are hosted under the OpenAI organization, spanning benchmarks, professional evaluations, and research packages. Noteworthy entries include genebench-pro-public-package (10 updates, 1.41k downloads), healthbench-professional (525 downloads, 1.9k views), gsm8k (17.6k updates, 1.14M downloads), and graphwalks (1.15k updates, 1.96k views). The collection also features coval (20.5k downloads), mrcr (2.4k downloads), and several others updated through late 2025. Recent paper activity includes GPT-Red: Automated Red Teaming via Self-Play at Scale and the observation that "Reasoning Models Struggle to Control their Chains of Thought." Dataset updates have continued through June 2025, with healthbench and frontierscience both receiving refreshes in December of that year.
Spaces and Collections
Six spaces are currently featured under the OpenAI org, with three collections visible in the interface. The OpenAI Privacy Filter ZeroGPU demo stands out as a featured space, tagged with a shield icon and labeled "OpenAI Privacy Filter." Other running spaces include Whisper for audio transcription (533 active users), Point E (111 users), and Safety GPT-OSS 20B for policy-guided response generation (2.83k users). A seventh space, openai-detector, carries a "detect if text is generated by GPT-2" tagline but appears paused since January 2023.
Collections group models thematically. The gpt-oss-safeguard collection links the 120B and 20B safety variants; the main gpt-oss collection hosts the full suite of open-weight language models; and a third collection rounds out the grouping. Each collection appears under the "View 3 collections" section on the organization page.
Team and Organization Structure
The OpenAI organization card lists 144 team members, with activity counters showing recent joins: +110, +97, +76, +66, +46. These figures appear in the team members sidebar and suggest a growing operational footprint alongside the model releases. The organization description on the page header emphasizes the community aspect: "Welcome to the official Hugging Face organization for OpenAI's open models! 🤗"
Recent Activity and Future Directions
Model updates continue to be sorted by "Recently updated," with gpt-oss-safeguard-120b most recently refreshed on October 29, 2025, and gpt-oss-safeguard-20b on January 14. The dataset roster similarly trends toward recent activity, with healthbench and frontierscience showing December 2025 updates. The six featured spaces and three collections indicate an expanding footprint beyond raw model distribution, moving toward tools, detection, and privacy-enhanced workflows. As OpenAI continues to publish open-weight assets, the Hugging Face presence serves as both a distribution channel and a pulse point for community engagement, download trends, and collaborative development.
Source: https://huggingface.co/openai
