gpt-oss-120b

Overview. gpt-oss-120b is OpenAI's open-weight, Apache-2.0 licensed reasoning model, released in August 2025 as the company's first meaningful open-weight release since GPT-2. It's a Mixture-of-Experts architecture with 117B total parameters and 128 experts, but only 5.1B parameters activate per token via top-4 routing, giving it a 131,072-token context window at a fraction of the compute cost of a dense model that size. Developers reach for it when they want OpenAI-grade reasoning behavior with the option to self-host, rather than being limited to a closed API.

Key strengths. The model is reasoning-tuned: it produces a visible chain-of-thought before its final answer, with a configurable effort level (low, medium, or high) that lets a caller trade latency for depth. Its MoE design is built for efficient serving despite the large total parameter count, which fits interactive, chat-style Inference workloads rather than heavy batch-only pipelines. Early signal suggests relative strength on math- and reasoning-oriented tasks compared with general writing or agentic tool-use scenarios, though this hasn't yet been confirmed against a full benchmark suite and should be treated as directional rather than definitive.

When to use it. Use gpt-oss-120b when a task benefits from an explicit, inspectable reasoning trace — multi-step math, logic puzzles, or problems where you want to dial reasoning effort up or down. It's also the model to reach for when open weights matter: teams that need to self-host, fine-tune, or run Inference on their own infrastructure for compliance or control reasons can do so under its Apache-2.0 licence. It is not built for image, audio, or document inputs, so route those workloads elsewhere.

Limitations. The model is text-in, text-out only — no vision, audio, or native document understanding. Its training data cutoff is roughly mid-2024, so it won't have knowledge of more recent events. Chain-of-thought reasoning is mandatory rather than optional, which means even simple queries incur some reasoning overhead, and tool-calling support exists in the model's format but hasn't been confirmed as a fully wired capability on every served endpoint.

Sovereignty note. gpt-oss-120b is created by OpenAI, a US company, and the currently configured endpoint runs on a US-based inference provider, so standard CLOUD Act exposure applies to that hosted path. Because the model is released under an open Apache-2.0 licence, organizations that need Sovereign, EU-resident Inference can instead self-host the weights on infrastructure of their own choosing rather than relying solely on the vendor-hosted endpoint.