Open Weights◆ AI-generated · Sourced
Release of gpt-oss-120b / gpt-oss-20b and gpt-oss-safeguard: Apache 2.0 Licensed Open-Weight Reasoning Models
TL;DR
OpenAI releases gpt-oss-120b and gpt-oss-20b — two state-of-the-art open-weight language models under the Apache 2.0 license — outperforming similarly sized open models on reasoning benchmarks; alongside them, gpt-oss-safeguard-120b and gpt-oss-safeguard-20b are introduced as policy-conditioned safety variants trained for content labeling per specified policies.
Model Positioning and Licensing
- gpt-oss-120b and gpt-oss-20b are open-weight language models released by OpenAI under the Apache 2.0 license, supplemented by the gpt-oss usage policy.
- Designed as cost-efficient reasoning models, they achieve superior real-world performance on mathematical reasoning, code generation, and multi-step logical tasks compared to comparably sized open models (e.g., Llama-3-70b, Qwen2-72b).
Deployment and Capability Optimization
- Both models are architecture- and quantization-optimized for efficient inference on consumer hardware: gpt-oss-20b supports FP16 full-precision inference; gpt-oss-120b supports 4-bit AWQ quantized inference on a single RTX 4090.
- They natively support tool use with JSON Schema output and function-calling protocols, making them suitable for agent-based applications.
Safety-Enhanced Variant: gpt-oss-safeguard
- gpt-oss-safeguard-120b and gpt-oss-safeguard-20b are post-trained safety-specialized variants derived from their respective gpt-oss base models.
- Their core capability is policy-conditioned reasoning: given a formal policy document (e.g., content moderation guidelines), they perform structured content labeling (e.g., flagging for violence, fraud, or PII exposure).
- The gpt-oss-safeguard technical report presents baseline safety evaluations against the original gpt-oss models, demonstrating improvements in policy adherence, refusal rate, and false-positive labeling.