GPT-OSS 20B is OpenAI's open-weight model, released in August 2025 under the Apache 2.0 license. Here it runs in a version with the built-in refusals removed. It uses a mixture-of-experts design: about 21 billion parameters, of which about 3.6 billion work on each token. That keeps answers quick, and the weights fit in 16 GB of video memory.
In short
GPT-OSS 20B Uncensored runs on cards with 16 GB or more. The most affordable suitable card right now is the RTX A4000 16 GB from 13 ₽ per hour, billed by the minute.
Updated 13 October 2026
The model needs a card with 16 GB of video memory: the weights take about 14 GB, and the context and buffers take the rest. An RTX 4080 with 16 GB is the minimum. For more room, an RTX 4090 with 24 GB or an RTX 3090 leaves space for the cache.
Prices and availability come from the live catalog and update automatically.
GPT-OSS reasons before it answers, like other models with thinking. A simple question gets a quick reply, and a hard one takes longer but is usually more accurate. That suits code and logic tasks. One server with a 16 GB card runs the whole model. For heavier work there is GPT-OSS 120B, which needs an 80 GB card.
The original model is released under Apache 2.0. The version without refusals was built from the open weights by third-party developers: the refusals were removed and the knowledge stayed the same. It is not an official OpenAI build. If you want the model for a product, read the LLM solution page and check the terms on the version's own page.
The API lives at https://<id>-11434.run.gpu.nz/v1 and follows the OpenAI format. Your rental token goes in as the key. A JavaScript example with no extra libraries:
const res = await fetch("https://<id>-11434.run.gpu.nz/v1/chat/completions", {
method: "POST",
headers: { Authorization: "Bearer <token>", "Content-Type": "application/json" },
body: JSON.stringify({
model: "huihui_ai/gpt-oss-abliterated:20b",
messages: [{ role: "user", content: "Explain the difference between let and const in JavaScript." }],
}),
});
const data = await res.json();
console.log(data.choices[0].message.content);The context is 16,384 tokens, and the model stays in memory while the server runs. Short answers arrive in seconds; hard questions take longer. The table compares it with the larger version. To work out memory needs, see the VRAM guide for LLMs.
| Spec | GPT-OSS 20B | GPT-OSS 120B |
|---|---|---|
| Weights | about 14 GB | about 65 GB |
| Video memory | 16 GB and up | 80 GB |
| Context | 16,384 tokens | 32,768 tokens |
| Replies | faster | slower |