GPT-OSS

Description: OpenAI's first open-weight models with reasoning capabilities

Website: https://github.com/openai/gpt-oss

OpenAI released GPT-OSS in August 2025 as its first open-weight models. These are optimized for reasoning tasks and available under Apache 2.0 license.

Model Variants

• gpt-oss-120b: 117B parameters (5.1B active), runs on 80GB GPU, achieves nearly GPT-o4-mini performance

• gpt-oss-20b: 21B parameters (3.6B active), requires only 16GB memory, ideal for local devices and low latency

Features

• Strong reasoning and tool-use capabilities
• Full chain-of-thought explanations
• Configurable reasoning effort (low, medium, high)
• Function calling and structured outputs
• Training informed by OpenAI o3

Performance

Surpasses similarly sized open-source models and sometimes even proprietary models like GPT-4o in specialized benchmarks.

Installation

Runs with vLLM, Ollama, llama.cpp. Not via OpenAI API, but as a local download model.