Overview
Run GPT-OSS 120B on Crusoe. Fine-tune and deploy OpenAI’s open-weight reasoning model with adjustable reasoning effort, native tool use, and Apache 2.0.
Explore additional models
Frequently
asked questions
GPT-OSS 120B is the larger of OpenAI’s two open-weight models, a mixture-of-experts with about 117B total parameters and 5.1B active per token. It supports native tool use, structured output, chain-of-thought reasoning, and a 131,072-token context, and it ships under the Apache 2.0 license. On Crusoe it is available for Serverless Inference, Serverless Fine-Tuning, and Self-Serve Deployments. You can also request a Tailored Deployment for bespoke inference optimization and a dedicated endpoint.
GPT-OSS 120B has about 117B total parameters with 5.1B active per token, while GPT-OSS 20B has 21B total with about 3.6B active. Both share the same training recipe, response format, and reasoning-effort controls. 120B is the higher-quality model, and 20B is faster and lighter to serve. On Crusoe, 120B is available for Serverless Inference, Serverless Fine-Tuning, and Self-Serve Deployments, while 20B is only available for Serverless Fine-Tuning and Self-Serve Deployments.
No, running GPT-OSS 120B on Crusoe never sends prompts, outputs, or fine-tuning data to OpenAI. GPT-OSS is an open-weight model, so Crusoe serves it on Crusoe-operated infrastructure rather than routing requests through OpenAI’s API. Data handling is governed by Crusoe’s infrastructure and policies, not by OpenAI, which makes GPT-OSS a way to get OpenAI-style reasoning while keeping data off OpenAI’s servers.
OpenAI reports GPT-OSS 120B reaching near-parity with o4-mini on core reasoning benchmarks while running as open weights. It is the model to choose when you want OpenAI’s reasoning behavior on infrastructure you control, with adjustable effort to trade depth for speed per request.
Yes. Use Serverless Fine-Tuning to customize GPT-OSS 120B on your own data. Upload a JSONL or Parquet dataset, run a LoRA job with pre-configured defaults, then go live in one click using Self-Serve Deployments or download the weights under Apache 2.0. Self-Serve Deployments offer inference profiles optimized for throughput or responsiveness. GPT-OSS 120B is also available for pay-as-you-go Serverless Inference, or you can request a Tailored Deployment for bespoke inference optimization and a dedicated endpoint.
Yes, GPT-OSS 120B supports three reasoning-effort levels: low, medium, and high, set in the system prompt, with medium as the default. Higher effort produces longer internal reasoning and better results on hard problems, at the cost of latency and output tokens. Since Serverless Inference bills output tokens separately from input, effort is a per-request lever on both response time and token usage.
MemoryAlloy technology is a proprietary memory fabric in Crusoe’s inference engine that shares the KV cache across nodes, with cache-aware routing that sends each request to the node most likely to have its context ready. GPT-OSS agents resend the same system prompt, tool definitions, and history on every call, so each turn skips redundant prefill and starts generating sooner, and reused context is billed at the lower cached-input rate.


%204.png)
