OpenAI's open weight models are here: gpt-oss-120b and 20b

Welcome OpenAI’s new advanced open-weight reasoning models to customize for any use case and run anywhere.

  • Permissive license
  • Designed for agentic tasks
  • Deeply customizable
  • Access to the Full chain-of-thought

Try both models in the browser. The playground is available here!

Or, start building right away! Download from Hugging Face or view the code on GitHub

Need more info?
Read the research blog and check out the model system card.

We also have a brand new category dedicated to the open source models. Looking forward to engaging discussions, exciting projects and all the questions we can answer together!

Let’s build!

34 Likes

Also see:

OpenAI Cookbook gpt-oss topics

6 Likes

Just found out that both models work with the Responses interface.

5 Likes

Indeed, and also with Codex CLI! (Need to update your codex package)

6 Likes

Did you find pricing? From the model page it looks like the rate limits for all tiers is 0, so this is surprising that it’s callable from the API.

2 Likes

OpenAI’s not serving it up, though (at least not to me, though it seems some even get it in the playground (the separate domain?)). No pricing, no rate, no model:

HTTP error occurred: 400 - {
  "error": {
    "message": "The requested model 'gpt-oss-20b' does not exist.",
    "type": "invalid_request_error",
    "param": "model",
    "code": "model_not_found"
  }

}

What appears currently just looks like they used the model page template for giving information, without clearing the template or copied database entry of its endpoint:true values.

2 Likes

You can set up your own responses compatible API.

Here are some direct links that @edwinarbus shared earlier with specific examples.

But I will change the post above to use the term interface instead of API to be more clear.

Regarding the pricing if hosted by another provider I found OpenRouter asking for 0.15 per million input tokens and 0.6 to 0.75 million output tokens.

7 Likes

Wow guys, you are completely on top of all of this. Thank you so much.

2 Likes

Wow it’s working on ollama and result is amazing

5 Likes

Probably the smartest move since quite a while. Bravo!!! Excellent news.

Is the fine-tuning UI for those coming along in the bundle? Or that would be to good? Kidding.

Thanks guys, awesome as always. Really appreciated.

3 Likes

Loving this How to run gpt-oss locally with Ollama and great work adding the section on AgentsSDK Integration with LiteLLM, just in time for my lecture tomorrow. Assuming I can use this in a Google Collab notebook as well for demos

5 Likes

The 120B model is screaming on my 2022 Mac with 128G of RAM.
:exploding_head:

Was not expecting this … speechless.

6 Likes

gpt-oss20b is timing out for a content generation usecase locally - Content generation and email tool call using AgentSDK doesn’t work with gpt oss( anyone else got it to work ? )
litellm.exceptions.APIConnectionError: litellm.APIConnectionError: OllamaException - litellm.Timeout: Connection timed out. Timeout passed=600.0, time taken=600.129 seconds

meanwhile, chatCompletions for content generation worked great.

3 Likes

@kavitatipnis

Can you create a separate topic for this issue?
I think this is something more people are experiencing at the moment.

3 Likes

OMG! They come with a ready MCP Server already ready! :face_holding_back_tears: :heart_eyes: This is an excellent launch. Good job!

4 Likes

GPT-5 : Plus puissant ? Peut-être… Mais moins agréable à utiliser !

Message :

Fidèle utilisateur d’OpenAI depuis longtemps, j’ai toujours apprécié la fluidité et la précision des échanges. Mais depuis le passage à GPT-5, je constate un vrai changement : moins naturel dans les réponses, plus formaté, et parfois moins pertinent dans le ton.

Ce n’est pas un problème de performances brutes, c’est une question d’expérience utilisateur. Avant, on avait l’impression de parler à un assistant flexible et “humain” dans sa manière d’échanger. Aujourd’hui, c’est plus froid, plus contraint, et on perd cette proximité qui faisait toute la différence.

Je ne cherche pas à “changer de crèmerie” — aucun concurrent ne propose l’écosystème complet qu’offre OpenAI. Mais justement, c’est pour ça que j’exprime mon mécontentement ici : on veut garder ce qui faisait la force de GPT.

OpenAI, écoutez vos utilisateurs avant que la fidélité ne se transforme en frustration. On veut la puissance ET la personnalité.

4 posts were split to a new topic: Image input for oss models?

Did you have a GPU? I tried the gpu-oss-20b on Ollama and it crashed. Any idea how I can do it without a GPU?

Try a gguf model maybe?
(i have not test it yet. can not recommend any now…)

(maybe test this… lmstudio-community/gpt-oss-20b-GGUF · Hugging Face)