OpenAI releases gpt-oss-120b and gpt-oss-20b under Apache 2.0
Key Questions
What models did OpenAI release under Apache 2.0?
OpenAI released gpt-oss-120b, a 120B MoE model with 128k context focused on tool use, and gpt-oss-20b, a 21B MoE with 3.6B active parameters. Both are open-weight and available for community use.
What is Harness-1 and how does it perform?
Harness-1 is a 20B retrieval subagent trained via RL on gpt-oss-20b that achieves 0.730 curated recall, outperforming other open models. It demonstrates strong results in agentic retrieval tasks.
What does this release mean for the open-source AI movement?
The release provides major validation of open-weight models and is already seeing community adoption, such as running GPT-OSS 20B on OpenClaw. It highlights growing competition with proprietary systems.
OpenAI released two open-weight models: gpt-oss-120b (120B MoE, 128k context, tool-use focus) and gpt-oss-20b (21B MoE, 3.6B active) under Apache 2.0. Harness-1, a 20B retrieval subagent trained via RL on gpt-oss-20b, achieves 0.730 curated recall, beating all open models. Community already running GPT-OSS 20B on OpenClaw. Major validation of open-source AI movement.