GPT OSS 120B (Databricks, EU)

GPT OSS 120B (EU) is delivered through an EU-hosted routing path that emphasizes data residency and compliance, giving European teams a way to run a frontier-class reasoning model without sending prompts outside the region. The offering is positioned for complex analytical work, logical reasoning, and mathematical problem-solving, and pairs those strengths with JSON mode and streaming for production-grade applications. Because the underlying weights are openly shared, organizations can also self-host or audit the model, which matters for regulated environments that need transparency alongside performance.

At the architecture level, GPT-OSS 120B is an open-weight Mixture-of-Experts transformer with 116.8 billion total parameters but only 5.1 billion active per token across 36 layers, routing through the top 4 of 128 experts at each step. It uses grouped-query attention with rotary positional embeddings over a 128k-token context window, a combination that supports long-document reasoning while keeping inference efficient enough to run on a single 80GB GPU. OpenAI designed it to reach near-parity with o4-mini on core reasoning benchmarks and to deliver strong function calling and structured-output behavior, making it a practical fit for agentic pipelines that need reliable tool use on top of solid analytical depth.