This week Opeli presented two models of open weight language -Gpt-OS-120B and GPT-OS-20B. These new models have been designed to ensure powerful reasoning possibilities, flexible use of tools and adaptation of a wider audience at the level of programmers, all under the Apsesive Apache 2.0 license.
In contrast to the reserved GPT-4 or GPT-4O models hosted only on OPENAI cloud infrastructure, GPT-OSS models are available to anyone who can download and operate locally or via various implementation platforms, enabling lower delay, inference at the test at the sight and improved data control.
GPT-OS-120B and GPT-OS-20B models are designed to achieve good tasks that require reasoning, and at the same time others. The flagship model 120B contains 117 billion parameters and activates only 5.1 billion parameters on token thanks to the architecture of the expert mix (MOE), allowing the model to launch the model on one GPU processor by 80 GB. Meanwhile, the 20b version uses 3.6 billion active parameters for token, requiring only 16 GB of memory – ideal for consumer laptops and edge devices.
Both models support 128,000 context windows, chain reasoning (COT) at low, medium and high effort and structured output forms. They also integrate with the possibilities of using tools, such as performing Python code and searching for networks-uncommon to supply the agency's work flows.
Training with the help of the most advanced OpenAI techniques, including a high computer learning, supervised tuning and the process of equalizing after training, GPT-OSS models divide the development line with models of Serials about OpenAI (e.g. O3, O4-Mini).
The models are based on rotational positioning (ROPE), low local attention and grouped multiplayer attention to speed and performance balance. Before training, he focused on STEM, programming and general knowledge, with a toxate based on Superset used by GPT-4O, known as O200K_HARMONY-OPEN SOURCE.
Opeli emphasizes that security was fundamental in the development of these open models. The company filtered pre -training data to avoid exposure to high -risk topics (e.g. chemical, biological, nuclear domains) and used hierarchies of equalization and instructions to increase resistance to opposite hints.
To simulate the worst scenarios of improper use, OpenAi has enemically refined models in the field of sensitive, such as cyber security and biology. However, even with intentional attempts to “arma” models using their own training stack, the models did not reach high -risk levels as defined as an OpenAI readiness. Independent reviews confirmed these findings.
In addition, OpenAI began a red team challenge with a prize pool of USD 500,000 for a further surface of all innovative security security, encouraging the global AI community to cooperate in testing the stress of models.
The models are free of charge on hugging the face, quantified in MXFP4 for performance performance. Opeli also issued tools to apply in Pytorch, Apple Metal and provided Harmony format renderers in Python and Rust.
Implementation partners include the main platforms such as Azure, AWS, Hisging Face, Vercel, Ollam, Lama.cpp and many others. On the front of the equipment, cooperation with NVIDIA, AMD, CEERBras and GRAQ provides optimized support on all devices.
Microsoft also introduces optimized GPU local versions of GPT-ASS-20B into Windows via ONNX Runtime, available via Foundry Local and AI Toolkit for Visual Studio code.
Despite their capabilities, GPT-OSS models are only textual and do not have multimodal functions such as understanding of image or sound. Their hallucinations remain much higher than newer reserved models, with GPT-ASS-120B hallucination in 49% of PersonQA comparative response, compared to 16% for O1.
In the case of GPT-OSS, OpenAI again opens the door to a transparent, decentralized AI development on a large scale. By balancing powerful possibilities with architecture aware of safety, these models allow researchers, startups and programmers to explore, tun and introduce innovation in world -class language models.
