OpenAI offers open AI models but CIOs must assess the risks

Openaiclaims that the new models are more efficient and can be deployed on consumer hardware.

OpenAI claimed that its gpt oss 120b model achieves nearly-parity with Openai O4 Mini is able to run efficiently on an 80 GB GPU, and performs well on core reasoning benchmarks. It said that the gpt oss-20b model can deliver similar results to OpenAI O3-mini in common benchmarks, and run on edge devices using only 16 GB of RAM.

Nvidia, a graphics processor unit (GPU), said that OpenAI’s models were trained on Nvidia GB200 NVL72 systems, and could use Nvidia NIM Microservices. It said this would allow for easy deployment on any GPU accelerated infrastructure, with flexibility, privacy, and enterprise-grade security. Nvidia

said that software optimisations to the Nvidia Blackwell platforms, can enable the models to achieve 1.5 million tokens every second when run on Nvidia GB200 NVL72 system, for AI inference.

Amanda Brock is the CEO of OpenUK. She said: “The beauty in open source and openness for AI is that it caters to all sides of the global debate. It has the power of being a digital good, creating access for everyone, but commercially it allows the creation of standards and promotes adoption. Think Meta’s Llama open innovation model It allows for global reach and impact, especially in a world of geopolitical change.”

An open AI model has the main benefit that it’s not closed. This means that anyone can check it. This will help improve the quality of the model, remove bugs, and reduce bias when data used to train a model are not diverse enough. Open models allow businesses to fine-tune their LLMs to the way they run. CIOs must weigh the pros and cons of using an open AI over a proprietary model, especially when they are faced with significant operational costs.

Haritha Khandabattu is a senior director and analyst with Gartner. She said that open models, popularised in LLMs like Meta’s Llama by LLMs, are used primarily in regulated industries. She said that these industries are more inclined to experiment with the open models. “Depending on how and where you deploy the open models, it may also require significant infrastructure.”

Khandabattu stated that organisations are experimenting open models to retain control. Khandabattu has said that the total cost for deployment is “very expensive” based on the IT leaders with whom she has spoken. Open models require significant engineering and operational costs to run, customize and maintain.

She said that open models for AI applications, such as AI-based coding may not always match performance of proprietary models. She said that this can negatively affect organisations, such as a reduced level of overall employee or developer experience and slower performance times.

Khandabattu encouraged IT leaders to weigh the pros and cons open models that may provide the level of enterprise IT assistance needed by an organisation. “Like open source enterprise software, they come with their own risks,” she added.