GPT OSS 120B (OVHcloud)

GPT OSS 120B is an open-weight Mixture-of-Experts language model from OpenAI that prioritises reasoning, agentic workflows, and general-purpose production workloads. The model carries 117B total parameters but activates only 5.1B per forward pass, using native MXFP4 quantization so it can run efficiently on a single high-end GPU. Training drew on a combination of reinforcement learning and techniques borrowed from OpenAI's more advanced internal systems, including o3, which the release credits for pushing reasoning ability into an open-weight release under the Apache 2.0 license.

In practice the model offers configurable reasoning depth, full access to chain-of-thought traces, and native tool use spanning function calling, browsing, and structured output generation. OpenAI reports that it reaches near-parity with o4-mini on core reasoning benchmarks and scores strongly on agentic evaluations, including the Tau-Bench suite and HealthBench, where it is described as outperforming proprietary peers such as o1 and GPT-4o. That combination of flexible reasoning controls, agentic capability, and a context window suited to long documents makes it a practical fit for teams that want an open, locally deployable model for multi-step problem solving and tool-driven applications.