GPT OSS 20B (Databricks)

GPT OSS 20B is a 20-billion-parameter open-weight model from the gpt-oss family that focuses on efficient chain-of-thought reasoning rather than sheer scale. It uses a compact mixture-of-experts design with SwiGLU activations and a token-choice routing scheme that is optimized for single-GPU deployment, paired with native FP4 quantization and a 128K context window. An alternating attention mechanism that mixes full and sliding-window contexts, along with a learned attention-sink layout, helps the model keep memory use manageable while still reasoning over long inputs, which makes it attractive for organizations that want frontier-style reasoning without frontier-scale hardware.

The model is positioned for practical agent and enterprise workloads: it supports tool use and chain-of-thought reasoning, and it is well suited to building fast agents, analyzing large document collections, and creating domain-specific custom assistants. Its 131K-token context, open-weight licensing, and compatibility with the standard GPT-4o tokenizer and Harmony format make it straightforward to fine-tune, deploy behind a governed platform, and integrate into retrieval-augmented or coding-agent pipelines where controllable, inspectable reasoning is more valuable than the largest possible model.