OpenAI: the GPT-OSS family of open models
Experiment with OpenAI GPT-OSS reasoning models, ready to customize and deploy on Crusoe infrastructure.
Get faster turns for agents and reasoning loops
Agentic workflows repeat system prompts, tool definitions, and history every call. Crusoe’s inference engine, powered by MemoryAlloy™ technology, caches shared context cluster-wide, allowing GPT-OSS agents to generate output quickly without reprocessing prior context.
Swap models without rewriting your code
Crusoe serves GPT-OSS models through OpenAI-compatible endpoints, so apps built on the OpenAI SDK can switch over by changing the base URL and API key. Keep your existing integration and run the open weights on Crusoe infrastructure.
Pricing that fits the workload
Pay per token while you experiment, then move to Self-Serve Deployments for predictable GPU-hour cost at production volume. Pair that with GPT-OSS’s adjustable reasoning effort to control latency and output tokens request by request.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
OpenAI is the San Francisco AI lab behind ChatGPT. In August 2025 it released its first open-weight models since 2019, the GPT-OSS family, under Apache 2.0, bringing its reasoning training and configurable reasoning effort to weights you can run outside OpenAI’s API. On Crusoe Cloud, GPT-OSS models are available across Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning, with support varying by model.
Yes, GPT-OSS 120B and GPT-OSS 20B are released under the Apache 2.0 license with no royalties, and you can fine-tune and redistribute them without open-sourcing your own changes. OpenAI also applies a separate usage policy alongside the license. On Crusoe, GPT-OSS 120B is available for Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning. GPT-OSS 20B is available for Serverless Fine-Tuning and Self-Serve Deployments.
Choose the smaller GPT-OSS model for low-latency, high-volume, or cost-sensitive workloads, and for fast fine-tuning iteration. Choose the larger model for complex reasoning, advanced coding, and agentic tasks where accuracy matters more than speed. Both use a mixture-of-experts architecture, so only a fraction of parameters activate per token. Both also support configurable reasoning effort, so you can tune speed and depth within either model.
GPT-OSS supports three reasoning effort levels — low, medium, and high — set through the system prompt, with medium as the default.. Higher effort produces longer reasoning and better results on hard problems at the cost of latency and output tokens. Since Serverless Inference bills output tokens separately from input, switching effort per request is a direct lever on both response time and token usage.
Yes, both GPT-OSS models are available for Serverless Fine-Tuning on Crusoe. The 20B model is the lighter option for fast iteration, while the 120B model gives you the larger base. Under Apache 2.0, you can export your tuned weights and deploy them anywhere, or serve them on a dedicated endpoint through Self-Serve Deployments.
GPT-OSS models are open-weight models you can download, run, and fine-tune on your own infrastructure. ChatGPT and OpenAI’s API models are proprietary and run only on OpenAI’s servers. GPT-OSS shares OpenAI’s reasoning training approach but is released at smaller sizes. Running GPT-OSS on Crusoe Cloud means prompts and outputs never pass through OpenAI.