Start chatting instantly with our ChatGPT free alternative. Our AI chatbot is a free AI assistant powered by GPT-5. You can ask questions, write emails, summarize PDFs, and create images easily. No login needed.
GPT-OSS is the model OpenAI published the weights for. That sentence would have been strange a couple of years ago, and it is the whole reason this page is interesting: the company that made closed models the norm released one you can download, inspect and run on your own machines. The 120B is the larger of the two sizes. On this page you can try it free, with no account, before deciding whether to host it.
Not because it beats GPT 5. It does not, and it is not meant to. It matters because it removes the dependency. A model you hold the weights to cannot be repriced, deprecated, silently updated, rate limited or taken away, and it can run somewhere no request is allowed to leave.
GPT-OSS 120B is OpenAI's open-weight release, published under a permissive licence and available to download and run yourself. It is a text model: it does not read images. It reasons noticeably better than its size would suggest, which is the thing people remark on when they first try it, and it has a knowledge cutoff like every model, so anything recent has to be given to it rather than recalled.
| Self-hosted | A hosted API | AIFreeForever GPT-OSS | |
|---|---|---|---|
| Price | Your hardware | Metered per token | Free |
| Setup | Download, GPUs, serving | Account and key | None |
| Data leaves your machine | No | Yes | Yes |
| Model can change under you | No | Possibly | Possibly |
| Good for | Production | Scale | Deciding |
That table says the one thing this page cannot do. If the reason you want an open model is that nothing may leave your building, this is a trial and not a solution. Use it to decide, then download the weights.
| Model | Weights | Reads images | Best at |
|---|---|---|---|
| GPT-OSS 120B | Open | No | Reasoning without a vendor |
| DeepSeek V4 | Open | No | Code and math |
| Meta Llama 3.1 | Open | No | Small and self-hostable |
| Microsoft Phi-4 | Open (MIT) | No | Maths for its size |
| GPT 5 | Closed | Yes | Hard reasoning, tool use |
This is the comparison the page exists for, and the honest answer is that GPT 5 is better. The question worth asking is by how much, on your work. Run the same ten real prompts through both from the model bar. For a lot of tasks the gap turns out to be smaller than the difference between a good prompt and a lazy one, and that is when self-hosting starts to look sensible.
Both are open and both are large. DeepSeek V4 is the specialist for code and mathematics and takes far more text in one go. GPT-OSS is the better generalist reasoner. If you are choosing one to host, choose by what your actual workload looks like, not by benchmarks.
Benchmarks and riddles tell you nothing about whether a model can handle your support tickets. Bring ten real examples.
Not a casual question. The prompt that would be in the code: explicit format, an example of a good answer, and a rule for what to do with unexpected input. That is the version whose quality you are actually measuring.
Consistency matters more than peak quality for anything automated. Ask the same thing twice and see whether the answers differ in shape.
Open GPT 5 in the bar and run it again. The gap you measure is the price of independence, in the only units that matter.
The chat box at the top of this page is already on GPT-OSS.
Give it a real task, with an example of the output you want.
Attach a document if you have one. PDF, Word and Excel are read directly; images are not.
Run the same prompt on GPT 5 from the bar and compare. That comparison is the point of the page.
Requests go over HTTPS through our server and conversations are not stored; uploaded files go at the end of the session. The honest caveat above still stands: this is a hosted service, so it does not answer the "nothing leaves my network" requirement. Downloading the weights does, and that is what they are published for.
It is a text model, so no images. It will state a specific fact it half-remembers rather than admitting the gap, which is the standard failure of every model and slightly more common outside the frontier tier. Its knowledge stops at its cutoff. And it is not GPT 5: on the hardest questions the difference is real and you will see it. For everything short of that, which is most things, it is a serious model that happens to be yours to keep.
Yes. No login, no card, no message cap.
No. ChatGPT runs OpenAI's closed models. This is their open-weight release, which anyone can download and run.
Yes. That is the point of it.
The 120B, the larger of the two released sizes.
No. For images use Gemini 3, Claude Haiku 4.5 or GPT-5 Mini from the bar.
No. Use Image mode on this page or the image generators in the menu.
Yes. PDF, Word and Excel.
No. This is an independent site offering free access to an openly published model.
Find out what an open model actually costs you in quality. GPT-OSS is free at the top of this page.
