OpenAI: gpt-oss-120b

gpt-oss-120b is an open-weight, 117-billion-parameter Mixture-of-Experts (MoE) language model from OpenAI, designed for high-reasoning, agentic and general-purpose production use. It is one of OpenAI’s first open-weight releases since GPT-2, distributed under the permissive Apache 2.0 licence, which allows commercial use, fine-tuning and self-hosting without restrictive terms.

The model activates roughly 5.1 billion parameters per forward pass and is optimised to run on a single 80GB H100 GPU using native MXFP4 quantization. That engineering choice is the headline: it brings near-frontier reasoning within reach of teams that can provision one data-centre GPU, rather than requiring a large multi-GPU cluster. It ships with a 131,000-token context window.

The Mixture-of-Experts design routes each token through a small subset of specialised expert sub-networks, so the model has the knowledge capacity of 117B parameters but the per-token compute cost of a roughly 5B model. This is what makes single-GPU deployment feasible while retaining strong reasoning. MXFP4 quantization compresses the weights further without the large quality loss associated with naive low-bit quantization.

gpt-oss-120b exposes configurable reasoning depth — you can trade latency for accuracy per request — and full chain-of-thought access, which is valuable for debugging, evaluation and building transparent agent systems. It supports native tool use including function calling, browsing and structured output generation, making it a genuine agent-capable model rather than a chat-only release.

On OpenAI’s published evaluations, gpt-oss-120b approaches the proprietary o4-mini on core reasoning benchmarks — a notable result for a model you can download and run yourself. It performs strongly on competition mathematics, graduate-level science and tool-using tasks, and is competitive on coding evaluations, particularly when allowed to use tools and higher reasoning effort.

As always, open-weight benchmark numbers depend heavily on the serving stack, quantization and reasoning settings you choose. The figures below are directional; the practical takeaway is that gpt-oss-120b closes much of the gap to small proprietary reasoning models while giving you full control over deployment, data handling and fine-tuning.

Choose gpt-oss-120b when control matters: data-sensitive environments that cannot send traffic to a hosted API, workloads that benefit from fine-tuning on proprietary data, or cost structures where self-hosting on owned or reserved GPUs beats per-token pricing at scale. The Apache 2.0 licence removes most legal friction around commercial deployment and derivative models.

It is also a strong choice for agentic systems that need transparent chain-of-thought and native tool use without vendor lock-in. The trade-off versus full GPT-5 is raw capability ceiling and the operational burden of running your own inference — you take on serving, scaling and reliability in exchange for control and potentially lower marginal cost.