Ā© Unsplash
Driven by increasingly fierce Chinese competition, OpenAI publishes its first open-weight GPT models since 2019. A more significant change than it appears: these models can run locally, without an internet connection, and are customizable as needed.
After having more or less let open source wither since the release of GPT-2 in 2019, OpenAI returns to its first loves. The new "open-weight" (or open-weight) models, gpt-oss-120b and gpt-oss-20b, are freely downloadable and can be run locally on personal or professional machines.
OpenAI is bringing out its heavy artillery
The largest model (120 billion parameters) requires a graphics card with 80 GB of memory, while the smallest (20 billion) can run with only 16 GB. The latter can therefore operate on some slightly more powerful consumer computers, including those equipped with Apple Silicon chips or modern graphics cards.
Users can freely customize these models, integrate them into their applications, and redistribute them, including in a commercial context, thanks to their Apache 2.0 license. After several years of keeping its proprietary models (GPT-3, GPT-4, GPT-4o) closed, OpenAI is therefore beginning to shift ā at least partially ā towards greater openness.
OpenAI insists, however, that gpt-oss is not a competitor to ChatGPT, nor to its API. These models are intended as a complementary offering, useful for researchers, developers, or companies wishing to run AI locally, including in sensitive or disconnected environments (hospitals, embedded systems, etc.).
The gpt-oss models incorporate reasoning techniques similar to those of the proprietary o3-mini and o4-mini models, which they largely outperform. In mathematics, healthcare, or software tool usage, performance is guaranteed, promises OpenAI. It is possible to use them as agents capable of executing code, browsing the web, or manipulating software.
In terms of security, OpenAI states that they have conducted extensive testing and no critical risks have been identified. Many technical partners (Microsoft, Hugging Face, AWS, Snowflake, Nvidia, etc.) have joined OpenAI to facilitate the deployment of the models in various environments, from the cloud to local devices. A toolset and guides are made available, as well as a testing playground.
It's not yet GPT-5, but the gpt-oss models already meet many needs: reduced cost, local execution, customization, and full control over data. According to OpenAI, these tools will help overcome barriers for emerging markets, resource-constrained sectors, and small organizations that do not always have the means to use proprietary models in the cloud.
« Wider access to these high-performing open-weight models, designed in the United States, contributes to embedding artificial intelligence in a democratic approach », adds the company, which does not forget to remind what is at stake: ensuring American dominance in global AI.
šš» Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.
Source : OpenAI
- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR
- Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Free
- ChatGPT
- Microsoft 365
- Opera One
- Avast One










- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Free
- ChatGPT
- Microsoft 365
- Opera One
- Avast One




