OpenAI gpt-oss

gpt-oss is OpenAI's return to open weights after the proprietary GPT-3/GPT-4 era: two mixture-of-experts reasoning models released August 5, 2025 under Apache 2.0 (plus a complementary use policy). The 120B model totals ~117B parameters with only ~5.1B active per token (128 experts, 4 active), runs on a single 80GB GPU, and supports a 128K-token context; the 20B variant (~21B total, ~3.6B active, 32 experts) is built for consumer hardware with as little as 16GB of memory. Both ship in MXFP4 quantization with three configurable reasoning-effort levels, full chain-of-thought access, instruction following, and tool use.

OpenAI positions the 120B model at or above its proprietary o4-mini and the 20B as rivaling o3-mini — provider-reported comparisons, not independent evals. The release drew a large fine-tuning and self-hosting community (Axolotl guides, Ollama/vLLM support, third-party hosted APIs), but more than a year later OpenAI has shipped no general-purpose successor, so the family is effectively a one-shot release. For builders who want OpenAI-descended reasoning weights they can run offline, it remains the only option.

• OpenAI-reported: gpt-oss-120b performs at or above o4-mini; gpt-oss-20b rivals o3-mini (provider-reported, not independently replicated)
• No canonical independent benchmark table verified; community evals tracked in third-party cookbooks