OpenAI launches GPT OSS: first open-weight model since GPT-2
OpenAI has made a big announcement: the company has released its first open‑weight models since GPT‑2 in 2019. These models, called gpt‑oss‑120b and gpt‑oss‑20b, are now freely available for download under the Apache 2.0 license.
Developers and researchers can run them locally, inspect their internal weights, customize, and even fine‑tune them for specific tasks.
What are GPT‑OSS models?
Open‑weight means that the model parameters are publicly accessible. The gpt‑oss‑20b variant is a 21 billion‑parameter model that can run on a desktop or laptop with around 16 GB of memory.
The larger gpt‑oss‑120b, with about 117 billion parameters, uses a Mixture‑of‑Experts (MoE) architecture and requires more powerful hardware like an Nvidia H100 GPU with about 80 GB of memory.
Both models support chain‑of‑thought reasoning, meaning they can think through multi‑step logic processes. They do not support images or multimodal capabilities, but they are designed for text tasks, coding, browsing, and even acting as agents integrated with other OpenAI models.
Why this release matters
This move signals a return to OpenAI’s roots. After years of primarily closed, proprietary launches, the company is re‑embracing transparency. The open‑weight release offers greater control for developers who care about data privacy, customization, and being able to run advanced models offline.
It also positions OpenAI more clearly in the open‑source AI ecosystem alongside Meta’s LLaMA, Mistral AI products, and Chinese offerings like DeepSeek.
Competitors like Meta and DeepSeek have already released successful open models, and OpenAI’s entry aims to reclaim leadership in this space. The release also coincides with its upcoming GPT‑5 launch expected in August 2025.
Rysysth insights
On the other side, what does this mean for developers, researchers, and organizations at large? For starters, the ability to run an OpenAI‑branded model offline is a big deal for anyone concerned about privacy, latency, or recurring cloud costs.
Because these models are open under Apache 2.0, they empower startups and academic labs to build customized tools without licensing barriers.
That can spark innovation in areas from healthcare to local language tools.
Although these models aren't as capable as OpenAI’s top‑tier proprietary models with image and voice features, they are solid text generators and thinkers. They benchmark close to o4‑mini in many tasks, especially coding and reasoning, but they are still text‑only and quite large for casual use.
OpenAI has also emphasized safety: the models underwent internal testing, including simulated misuse scenarios, and passed their Preparedness Framework evaluations. So despite being open, they were designed with safeguards in mind.
What comes next
It’s likely that OpenAI will continue developing proprietary models in parallel, including GPT‑5. But now the community also has real access to parts of its technology stack. Moving forward, we can expect more open collaboration, dataset sharing, and creative uses of these models.
If you are building AI systems or teaching students about language models, this launch opens new possibilities. You can now explore how a top‑tier GPT‑style model works under the hood, run experiments locally, and contribute your own improvements. It’s a step toward more equitable and open innovation in AI.
Until next time.