OpenAI releases gpt-oss, open-weight reasoning models
gpt-oss-120b (117B total, 5.1B active; fits on one 80 GB H100) and gpt-oss-20b (21B total, 3.6B active) are mixture-of-experts reasoning models released under Apache 2.0 with low/medium/high…
- Date
- 5 August 2025
- Who
- OpenAI
- Confidence
- High
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): OpenAI · Confidence: High gpt-oss-120b (117B total, 5.1B active; fits on one 80 GB H100) and gpt-oss-20b (21B total, 3.6B active) are mixture-of-experts reasoning models released under Apache 2.0 with low/medium/high reasoning effort and tool use (model listings); the model card says they were trained with large-scale distillation and RL and that the CoT was not directly supervised, so researchers can study it (arXiv 2508.10925, v1 2025-08-08). I infer that it is OpenAI's response to R1-era pressure. Sam Altman had said on 2025-01-31 that OpenAI had been "on the wrong side of history" on open weights (B08-16), and on 2025-03-31 OpenAI announced plans for its first open-weight model since GPT-2, with reasoning comparable to o3-mini (TechCrunch); the release came 197 days after R1. For open-weights context see B19.