OpenAI releases first open source models in 6 years amid China competition

OpenAI has returned to its open source origins with the release of two new frontier language models: gpt-oss-120b (120 billion parameters) and gpt-oss-20b (20 billion parameters). This marks the company’s first open source language model release in over six years, positioning OpenAI to compete directly with the surge of high-performing open source models from Chinese competitors like DeepSeek while offering enterprises maximum privacy and control over their AI deployments.

The big picture: OpenAI’s strategic pivot back to open source reflects mounting competitive pressure from Chinese AI companies that have released powerful open source models matching proprietary performance at zero cost.

• Business appears robust with $13 billion in annual recurring revenue and 700 million weekly ChatGPT users, but the majority of OpenAI’s API customers are already using a mix of paid OpenAI models and competing open source alternatives.
• The company aims to become a “full-service, full-stack, one-stop shop” for all AI needs, spanning proprietary services, APIs, and now open source offerings.

Key capabilities: Both models deliver frontier-level performance while running entirely offline for maximum privacy.

• The gpt-oss-120b runs on a single Nvidia H100 GPU and matches or exceeds OpenAI’s proprietary o4-mini model on reasoning and tool-use benchmarks.
• The smaller gpt-oss-20b can run locally on consumer laptops and is comparable to o3-mini, even surpassing it in some evaluations.
• Both support 128,000 token context length (roughly 300-400 pages of text) and are multilingual, though OpenAI declined to specify which languages.

Licensing advantage: The Apache 2.0 license offers unprecedented freedom for enterprise deployment without usage restrictions.

• Unlike Meta’s Llama license, which requires paid licensing for services exceeding 700 million monthly active users, OpenAI’s models have no such limitations.
• Enterprises can download, modify, and monetize the models without paying OpenAI anything, while maintaining complete data privacy by running them on internal hardware.
• For regulated industries like finance, healthcare, and government, this eliminates the risk of data being subpoenaed from cloud providers.

Technical architecture: Both models use a Mixture-of-Experts (MoE) design optimized for efficiency and reasoning.

• The gpt-oss-120b activates 5.1 billion parameters per token out of 117 billion total, while gpt-oss-20b activates 3.6 billion out of 21 billion total.
• Models support chain-of-thought reasoning, tool use, and few-shot function calling, with compatibility for OpenAI’s Responses API.
• The “o200k_harmony” tokenizer is also being open-sourced, giving developers complete control over the inference pipeline.

In plain English: Think of these models like having multiple expert consultants on call, but only activating the specific experts needed for each question to save computing power. Instead of using all 120 billion “brain cells” at once, the larger model only uses about 5 billion for each response, making it much more efficient while still delivering sophisticated answers.

Safety measures: OpenAI conducted extensive adversarial testing to ensure the models don’t pose frontier-level risks.

• The company filtered chemical, biological, radiological, and nuclear threat data during training and applied advanced post-training safety methods.
• Malicious fine-tuning scenarios—some of the most sophisticated evaluations to date—showed even weaponized versions remained below “High” capability thresholds for biorisk and cybersecurity.
• Three independent expert groups reviewed the methodology, and OpenAI partnered with SecureBio, a biosafety research organization, for external biology-focused evaluations.

Market competition: OpenAI enters an increasingly crowded open source landscape dominated by Chinese models.

• DeepSeek-R1, Qwen 3, and GLM-4.5 all offer Apache 2.0 licensing with performance approaching proprietary models.
• European competitors include Mistral’s Mixtral series, while U.S. alternatives like Meta’s Llama and Google’s Gemma carry more restrictive licensing terms.
• Hugging Face download data shows Chinese open source models like Qwen2.5-7B-Instruct leading adoption among developers.

Business implications: The release raises fundamental questions about the future economics of AI.

• As CEO Sam Altman previously suggested AI might become “too cheap to meter,” the proliferation of high-quality open source models threatens traditional subscription and API revenue streams.
• OpenAI is reportedly offering in-house engineers to help enterprise customers customize deployments, similar to Palantir’s “forward deployed” model.
• The company faces the challenge of justifying sky-high valuations while potentially cannibalizing its own paid services.

Availability and ecosystem: The models are immediately available across major cloud platforms and hardware providers.

• Downloads are live on Hugging Face and GitHub, with deployment support from Azure, AWS, Databricks, Cloudflare, and others.
• Hardware partners include Nvidia, AMD, and Cerebras, with Microsoft providing GPU-optimized Windows builds via ONNX Runtime.
• OpenAI announced a $500,000 Red Teaming Challenge on Kaggle to identify potential misuse pathways and accelerate safety research.

What they’re saying: OpenAI executives framed the release as returning to the company’s foundational mission while acknowledging competitive realities.

• “This is the first time we’re releasing an open-weight language model in a long time… We view this as complementary to our other products,” said co-founder and president Greg Brockman.
• The move comes after CEO Sam Altman expressed regret about being on the “wrong side of history” regarding open source releases and committed to new open models in March.