Table of contents
OpenAI announced last night the release of two open-weight language models, gpt-oss-120B and gpt-oss-20B, available under the Apache 2.0 license. This move marks a turning point for the company, which had not offered an open-weight LLM since GPT-2. The model weights are publicly accessible on Hugging Face.
Models Designed for Reasoning and Efficiency
The two models are based on a Mixture-of-Experts (MoE) architecture, with 117 billion and 21 billion parameters in total, respectively, but activating only a fraction (5.1B for the 120B, 3.6B for the 20B) for each token. Both support an extended context length of up to 128,000 tokens.
OpenAI claims competitive performance on reasoning tasks. GPT-OSS-120B would achieve results close to o4-mini on standard benchmarks (MMLU, HLE, TauBench...), while being executable on a single 80 GB GPU. The lighter 20B model is said to operate with 16 GB of memory, making it potentially usable locally or on embedded devices.
Compatibility and Use Cases
These models are compatible with OpenAI's Responses API and natively support Chain-of-Thought (CoT), function calls, structured outputs, and reasoning effort adjustment according to the task.
OpenAI targets usage in agentic workflows, intelligent assistant development, research, or local deployment for security or data sovereignty reasons. Partners like AI Sweden, Orange, and Snowflake were involved prior to the launch to explore concrete integration cases.
Security and Risk Assessment
OpenAI has long explained its shift towards closed models due to security issues. Security has thus been at the core of the company's considerations, leading to several postponements of this much-anticipated Open Weight model release. OpenAI now claims to have integrated advanced filtering and post-training mechanisms to mitigate the risks associated with public availability. An evaluation by external experts was conducted, notably on deliberately fine-tuned malicious versions (cybersecurity, biology), as part of OpenAI's Preparedness Framework.
According to the company, even in these extreme scenarios, the models would not reach concerning capacity levels. A $500,000 red teaming challenge has also been launched on Kaggle to encourage collaborative vulnerability detection.
A Controlled Return to Open Source?
This launch raises several questions. On one hand, it demonstrates a desire to rebalance the offering between powerful proprietary models and open source alternatives. On the other hand, it allows OpenAI to maintain a technical edge while framing usage by setting new security standards for open-weight.
The publication of weights under a permissive license, the tools provided (optimized inferences, harmony renderer, PyTorch and Metal support...), as well as partnerships with Azure, Hugging Face, or Vercel aim to facilitate adoption in an increasingly fragmented ecosystem.
It remains to be seen to what extent these models will be adopted by the community, especially against alternatives like Mistral, LLaMA, Mixtral, or Yi, and whether their effective openness (notably the possibility of free fine-tuning) will meet the expectations of researchers and developers.
Discover the model cards on Hugging Face:
https://huggingface.co/openai/gpt-oss-20b
https://huggingface.co/openai/gpt-oss-120b
Cet article publirédactionnel est publié dans le cadre d'une collaboration commerciale