OpenAI Releases Its First Open-Weight Models Since GPT-2 — gpt-oss-120b and gpt-oss-20b Under Apache 2.0
OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0 — the company's first open-weight models in years. The 120B model runs on a single 80GB GPU at near-o4-mini performance. The 20B fits on 16GB consumer hardware and matches o3-mini on key benchmarks.
Original sourceOpenAI released gpt-oss-120b and gpt-oss-20b, marking the company's first open-weight model release since GPT-2. Both models are licensed under Apache 2.0, enabling unrestricted commercial use and fine-tuning without royalties or platform restrictions.
The 120B parameter model achieves near-parity with OpenAI's own o4-mini on core reasoning benchmarks while running on a single 80GB GPU — making it deployable on infrastructure most organizations already have. The 20B model matches o3-mini on common benchmarks and runs on consumer hardware with just 16GB of RAM, opening the door for local deployment across a wide range of use cases.
Both models were trained using reinforcement learning techniques informed by OpenAI's frontier models including o3, and demonstrate strong tool-use capabilities designed for agentic applications. The models are available on Hugging Face, Microsoft Azure AI Foundry, Databricks, and can be locally accelerated on NVIDIA GeForce and Quadro hardware.
The move is widely interpreted as a strategic response to Meta's Llama 4 and Alibaba's Qwen 3.6, which have dominated the open-weight leaderboards. By entering the open-weight market with Apache 2.0 licensing and performance competitive with its own commercial models, OpenAI is reclaiming relevance in a market segment it had ceded entirely.
Panel Takes
“Apache 2.0 + near-o4-mini performance is the combination that unlocks real commercial use. The legal clarity alone is worth more than incremental benchmark gains.”
“This is OpenAI doing damage control after watching Meta and Alibaba capture the open-weight developer market. The timing — releasing open weights when their closed models are multiple generations ahead — is strategic, not generous.”
“The open-weight frontier just got a major new competitor. When every major lab is racing on open weights, the baseline for what's freely available to developers improves faster than any single lab could manage alone.”