OpenAI: gpt-oss-20b

gpt-oss-20b is an open-weight, 21-billion-parameter language model released by OpenAI under the Apache 2.0 licence. It uses a Mixture-of-Experts (MoE) architecture that activates roughly 3.6 billion parameters per forward pass, optimised for lower-latency inference and deployability on consumer or single-GPU hardware — including machines with around 16GB of memory.

It is the smaller counterpart to gpt-oss-120b, sharing the same open-weight, agent-capable design philosophy but at a size that runs comfortably on a workstation, edge device or laptop-class GPU. It ships with a 131,000-token context window and is trained in OpenAI’s Harmony response format, which structures reasoning and tool calls.

Like its larger sibling, gpt-oss-20b uses a Mixture-of-Experts design so that only a fraction of its parameters are active per token — about 3.6B of 21B — keeping inference fast and memory-efficient. This is what allows it to run on hardware that could never host a dense model of comparable quality, opening up genuinely local and offline use cases.

The model supports configurable reasoning levels, fine-tuning, and agentic capabilities including function calling, tool use and structured outputs. Training in the Harmony response format gives it a consistent structure for separating reasoning from final answers and for invoking tools, which simplifies building reliable agents on top of it.

On OpenAI’s published evaluations, gpt-oss-20b approaches the proprietary o3-mini on core reasoning benchmarks — impressive for a model that runs on consumer hardware. It is competitive on mathematics, science and tool-using tasks, particularly when given higher reasoning effort, and is well suited to being fine-tuned on narrow domains where a small specialised model can match much larger generalists.

As with any open-weight model, real-world results depend on your quantization, serving stack and reasoning settings. The directional figures below illustrate the point: gpt-oss-20b delivers a large share of small-frontier-model capability at a footprint you can run locally, which is its core value proposition.

Reach for gpt-oss-20b when you need local, offline or edge inference: privacy-critical applications, on-device assistants, air-gapped environments, or development where you want to iterate without API costs. Its size makes it practical to run on a single workstation GPU, and the Apache 2.0 licence removes commercial friction.

It is also an excellent fine-tuning base. For a well-defined task, a fine-tuned gpt-oss-20b can rival much larger general models at a fraction of the serving cost. The trade-off is the lower capability ceiling versus gpt-oss-120b and the proprietary GPT-5 tiers — for the hardest open-ended reasoning you will want a larger model, but for focused, high-volume or privacy-bound work, 20b is often the sweet spot.