Article complete

OpenAI Releases Two Open-Weight Models Ahead of GPT-5

The larger model, gpt-oss-120b, achieves near parity with OpenAI’s o4-mini on reasoning tasks and can run on a single 80 GB GPU.

Four glitchy busts wearing colorful sunglasses against a purple background, representing modern AI themes.

Key takeaways

  • OpenAI releases two new open-weight models, gpt-oss-120b and gpt-oss-20b, under Apache 2.0 license.
  • gpt-oss-120b achieves near parity with o4-mini on reasoning tasks, running on a single 80 GB GPU.
  • gpt-oss-20b is optimized for devices with 16 GB memory, similar performance to o3-mini.
  • Both models support configurable reasoning effort levels, allowing developers to manage latency and performance.
  • This launch follows China's release of advanced open-source AI models, intensifying global competition.

Takeaways drafted with AI from the article text.

https://twitter.com/sama/status/1952777539052814448

The larger model, gpt-oss-120b, achieves near parity with OpenAI’s o4-mini on reasoning tasks and can run on a single 80 GB GPU. The smaller gpt-oss-20b delivers similar results to o3-mini and is designed to operate on devices with just 16 GB of memory.

“These models are compatible with our Responses API⁠ and are designed to be used within agentic workflows with exceptional instruction following, tool use like web search or Python code execution, and reasoning capabilities,” OpenAI said in a statement. “They’re optimised for use in agentic workflows like tool use, web search, and code execution.”

Both models support configurable reasoning effort levels—low, medium, and high—giving developers control over latency and performance trade-offs.

This launch from OpenAI comes after China recently released several advanced open-source models, including DeepSeek V3, Kimi K2, Qwen3-Coder, GLM-4.5, and MiniMax-M1, signalling a major push in global AI competition.  Interestingly, all the top 10 trending models on Hugging Face are currently from Chinese AI labs.

https://twitter.com/natolambert/status/1952370970762871102

In benchmark evaluations, gpt-oss-120b surpassed OpenAI o3‑mini and performs at the level of or better than o4-mini in areas such as competitive coding (Codeforces), general problem-solving (MMLU and HLE), and function calling tasks (TauBench). It also delivers stronger results than o4-mini on health-focused benchmarks (HealthBench) and mathematical competitions (AIME 2024 and 2025). 

gpt-oss-20b also matched or outperformed o3-mini on the same benchmarks despite its smaller size.

The models use a mixture-of-experts architecture. gpt-oss-120b has 117 billion parameters in total but activates only 5.1 billion per token using 128 experts per layer. 

gpt-oss-20b activates 3.6 billion parameters per token with 32 experts. Both models support context lengths up to 128k tokens and employ techniques such as grouped multi-query attention and Rotary Positional Embeddings.

The training data primarily consisted of English text, with an emphasis on STEM, code, and general knowledge. The models were tokenised using the new o200k_harmony format, which has also been released publicly.

These are the first open-weight language models from OpenAI since GPT-2.The models were post-trained using techniques developed for the o4-mini series, including reinforcement learning and supervised fine-tuning.

On safety, OpenAI said it conducted adversarial fine-tuning to simulate malicious use cases and evaluated the outcomes using its Preparedness Framework. 

To encourage broader community participation in safety research, the AI company is launching a Red Teaming Challenge with a $500,000 prize fund. A report and evaluation dataset based on the challenge will be published later.

gpt-oss model weights are available on Hugging Face, with native quantisation in MXFP4 format for efficient memory use. Reference implementations are provided for PyTorch and Apple Metal, along with a harmony prompt renderer in Python and Rust.

Deployment partners at launch include Azure, Hugging Face, vLLM, Ollama, llama.cpp, LM Studio, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, and OpenRouter.“Microsoft is also bringing GPU-optimised versions of the gpt-oss-20b model to Windows devices,” OpenAI said 

The company said that while gpt-oss is ideal for developers seeking customisable, local deployment, its proprietary API models remain the best fit for multimodal support and integrated platform tools. API support for gpt-oss is under consideration based on developer feedback.

“Open models complement our hosted models,” OpenAI said, adding that it allows developers to innovate and build safely with AI on their own infrastructure, regardless of geography or resource availability.

Published Updated

Spotted an error? Tell our editors. Read our editorial standards and corrections.

More in this story

OpenAI · 5 stories
  1. Sam Altman Confirms OpenAI’s Project Strawberry
  2. You are here: OpenAI Releases Two Open-Weight Models Ahead of GPT-5
  3. OpenAI Appoints Key Leads in India as Altman Plans Delhi Trip Next Month
  4. Snowflake, OpenAI Ink $200 Mn Deal to Deploy AI Agents for Enterprises
  5. Google Gemini 4 Struggles With Coding Despite Strong Benchmark Numbers: Report
About the author

Tech Journalist

Siddharth is a media graduate who loves to explore tech through journalism and putting forward ideas worth pondering about in the era of artificial intelligence.

All stories
Belamy · Daily

The day’s AI news, before your first meeting.