GPT open source: what is actually downloadable, what open weights means, and how to choose one

The GPT models that power ChatGPT are not downloadable. When people search for "GPT open source" they are almost always looking for an open weight model: a model whose trained weights are published so you can download them and run the model on your own hardware or your own cloud account. OpenAI publishes two of its own. Its announcement of gpt-oss, dated 5 August 2025, released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 licence, and states that the 120b model runs on a single 80 GB GPU while the 20b model runs on devices with 16 GB of memory. Other families published for download include Llama, Mistral, Qwen, DeepSeek and Gemma. The decision that matters is not which name you pick first, it is reading the licence and sizing the hardware.

Open weights and open source are two different things

The Open Source Initiative, which maintains the Open Source Definition for software, has published an Open Source AI Definition setting out the principles and requirements for an AI system that grants users the freedoms to use, study, modify and share it. It also maintains a separate page on open weights, because the two terms describe different amounts of openness.

• Open weights. The trained parameters are downloadable. You can run the model, fine tune it and serve it. The training data and the full training code may or may not be published.
• Open source by the OSI definition. Enough is released, including information about the data and the code used to build the system, that somebody else could study and recreate it, with licence terms that permit use, study, modification and sharing.
• Open licence versus custom licence. Some model weights ship under a standard open licence such as Apache 2.0 or MIT. Others ship under a custom community licence with its own conditions, for example on acceptable use, attribution, or usage above a scale threshold. The Llama licence is published in full and is worth reading before you build on it.

Practical consequence: decide what you need the licence to allow before you shortlist models. Serving a model to paying customers, fine tuning on customer data, and redistributing a modified model are three different permissions, and they are the three clauses to read first.

What is actually available to download

These are the families most commonly deployed, named neutrally. Each publishes its own model cards and licence terms, which is where the current sizes and conditions live.

Family	Where the models and cards are published
gpt-oss	OpenAI's open models page and the gpt-oss-120b model card
Llama	llama.com downloads and the meta-llama collection
Mistral	mistral.ai and the mistralai collection
Qwen	the Qwen collection
DeepSeek	deepseek.com
Gemma	the Gemma model cards on Hugging Face

Model names and sizes in this category change every few months. Use the Hugging Face model index to see what is current, and read the model card rather than a round up, because the card carries the context length, the licence, the intended use and the evaluation the publisher stands behind.

What running your own model gets you

• Data stays where you put it. The strongest reason most organisations do this. Inference happens inside your network or your cloud account, which makes the data handling question answerable in a procurement review.
• Predictable cost at volume. You pay for hardware and operations rather than per token, which favours steady high volume workloads such as classification, extraction and summarising large archives.
• Specialisation. Fine tuning a mid sized open model on your own labelled examples gets very good results on one narrow task, often better than a general model for that task.
• Offline and edge use. A small model on a laptop or a device keeps working with no network, which matters for field work and for anything latency sensitive.
• Version control. The weights you downloaded are the weights you will have next quarter, so a result you validated stays reproducible.

Hosted frontier models remain the practical choice for the hardest open ended reasoning, for the newest tool and agent features, and for teams with no infrastructure capacity. Most organisations end up running both: an open model for the high volume narrow jobs, a hosted model for the hard ones. OpenAI publishes the current hosted line up in its models documentation.

How to choose one, by job

The job	What to look for in the model card
Classify or tag high volume text	A small model, 7B to 20B class, instruction tuned, cheap to serve, and good enough on your own test set
Extract structured data from documents	Reliable JSON or function calling output, a long context window, strong instruction following
Summarise long archives	Long context window and a licence that permits commercial use on customer content
Code assistance in a closed network	A code trained variant and permission to serve internally to employees
Fine tune on your own examples	Published fine tuning recipe, permissive licence for derivative weights, and a base variant as well as an instruct variant
Run on a laptop or a device	Small parameter count plus available quantised builds, and a stated memory requirement

Build a test set of fifty to two hundred real examples from your own work before you compare anything. Published benchmarks tell you which models are in the same league; your own test set tells you which one to ship.

Running one in practice

• To try it on your own machine, Ollama and LM Studio both pull a model and run it locally with a few commands or clicks.
• To serve it to an application, an inference server such as vLLM handles batching and throughput, and most expose an OpenAI compatible endpoint so your existing client code points at a new base URL.
• Quantisation reduces the memory a model needs by storing weights at lower precision. It is the normal way a large model fits on a smaller GPU, and the model card or the community build will say which quantisations are available.
• Budget for operations, not just hardware: monitoring, version pinning, an evaluation suite you rerun on every upgrade, and a rollback path.

This page covers downloadable models and how to choose between them. For how assistants built on these models decide which companies to name, see LLM optimisation, and for the terminology around it, how LLMO relates to GEO, AEO and SEO.

The GPT open source questions worth an answer

Is GPT open source?

The GPT models behind ChatGPT are available through OpenAI's API and products rather than as downloadable weights. OpenAI does publish open weight models under its gpt-oss name, released under the Apache 2.0 licence according to its announcement of 5 August 2025, and those you can download and run yourself.

What is the difference between an open source model and an open weight model?

Open weight means the trained parameters are published for download. Open source, in the sense the Open Source Initiative defines for AI systems, additionally covers the information needed to study and recreate the system and licence terms that permit use, study, modification and sharing. Many popular downloadable models are open weight, and the licence on each one is the thing to read.

Can I use an open model commercially?

That depends on the specific licence, which is published with every model. Apache 2.0 and MIT permit commercial use broadly. Custom community licences add conditions, which can include acceptable use terms, attribution and naming requirements, or terms that apply above a defined scale. Read the licence for the exact model and version you intend to deploy, and keep a copy with your deployment notes.

What hardware do I need?

Read the stated requirement on the model card and plan from there. OpenAI's gpt-oss announcement gives a single 80 GB GPU for the 120b model and 16 GB of memory for the 20b model, which is a useful scale reference: small models run on a good laptop, mid sized models run on one server GPU, and the largest open models need multiple GPUs or a managed host.

Will running my own model help my business appear in AI answers?

Those are separate jobs. Running your own model is about doing work inside your own systems. Appearing in the answers that ChatGPT, Gemini, Perplexity and Google AI Overviews give to buyers depends on what those products can find and cite about you on the open web, which is a publishing and reputation job. Do both, and keep them in separate plans.

The other side of the same technology

Whichever model you run internally, your buyers are using somebody else's. When they ask an assistant for the best option in your category, the answer names a few companies and it is assembled from what the assistant can find and cite about each one. That answer changes per prompt and it is not a ranking, so it needs its own prompt set and its own record of sources. AI Knows Us measures whether assistants name a business and where the answer came from, with a free scan and no account needed at aiknowsus.com.