GPT-OSS 120B on a single GPU: chat and API

What you can do

• hard multi-step problems: reasoning, mathematics, planning
• large technical texts and specifications, with up to 32,768 tokens of context
• code review: finding bugs in a module, refactoring ideas, architecture explanations
• a reference model for checking the answers of smaller ones
• an API for an internal team service that needs a strong model

Video memory

80 GB+

What opens

Ready chat and API

Which GPU you need

You need a card with 80 GB of video memory: the weights take about 65 GB, and the 32,768-token context needs most of the rest. An H100 SXM with 80 GB or an A100 SXM4 with 80 GB will do. 48 GB cards such as the L40S and RTX A6000 are too small, because the model will not fit in them whole. If a smaller model is enough, GPT-OSS 20B runs on 16 GB.

How to start

1. Open the catalog, pick an 80 GB card such as an H100 or A100, and press «Rent» with the GPT-OSS 120B template.
2. The server starts in 2–5 minutes, and the roughly 65 GB of weights download on first start: usually 10 to 30 minutes, sometimes longer.
3. When the model is ready, open the chat from the Open WebUI button or the Telegram link. The model is already selected.
4. Take the API address from the «Ollama API» button and the token from the server page, and use them in the examples below.
5. When you finish, stop the server. An 80 GB card bills noticeably more per hour than a 24 GB card, and stopping it ends the GPU charge at once.

Questions and answers

Other templates

• GPT-OSS 20B UncensoredOpenAI's open model with refusals removed. Fast, fits in 16 GB of VRAM.
• Qwen 3.8Alibaba's latest 27B model: code, writing, analysis, strong multilingual skills.
• Qwen 3.8 UncensoredChat with Qwen 3.8 27B, refusals removed. The model is preloaded: open the link and type.
• Hermes 4.3 36BNous Research's model for open conversation and role-play; it rarely refuses.
• Gemma 4 26BGoogle's model on a fast MoE architecture: reads images and long documents.
• Qwen3 14B UncensoredA light model with refusals removed, for budget GPUs from 12 GB.
• Open WebUI + OllamaYour own ChatGPT with any model from Ollama or Hugging Face: name it at launch or pull it in the chat.
• vLLMFast LLM inference with an OpenAI-compatible API for your own apps.

Useful guides

• Run LLMs on a rented GPU: Llama, Qwen, DeepSeek, MistralRun Llama, Qwen, DeepSeek and Mistral on a rented GPU: Ollama with a chat UI or vLLM with an API. Which card you need, what an hour costs, and how to start.
• How much VRAM an LLM needs: a table and a rule of thumbHow to estimate memory for model weights at fp16, int8 and int4, how much 8B, 14B, 32B and 70B models need, and why context also uses memory.
• Ollama and vLLM on a rented GPU: chat and API in 10 minutesStep by step: a model in Ollama with chat and API, and vLLM with an OpenAI-compatible server on a rented GPU. Commands, checks and common mistakes.
• Local AI models without a powerful computerWhat a local AI model is, how much video memory text, image and video models need, and how to run the same open models on a rented GPU.