ChatMDR and ChatIVDR 3.1 with Reasoning model

• Fast
  Designated by the rocket, this is the quickest model. It is great for concise answers that don’t require so much reasoning. The underlying model is Meta LLama 3 with 70 billion parameters. Example question for the fast model: Explain to the CEO of my company the difference between UDI-DI, UDI-PI and direct marking.
• Smart
  The smart model, with the lightbulb, is much better at more involved questions that require some constructive thinking. It runs on OpenAI GPT 4o.
  Example question for the smart model: What is the device classification of a linear accelerator for cancer radiotherapy?
  The smart model is also very good at follow-up questions. After ChatMDR returns the answer, you can for instance ask: Show me the steps how you came to that conclusion
• Reasoning
  This is the new model, based on GPT o1, more on which below. It does reasoning in steps and is therefore very good at breaking down a complex problem into multiple steps, which it will share with you. The reasoning model takes more time to do its work and tends to produce very verbose output. Please note: due to the high computation cost, the Reasoning model is only available to ChatMDR/ChatIVDR Plus and Enterprise customers.
  Example question that the reasoning model excels at:
  I have developed a medical device that I want to offer to a customer in France. Because my notified body is slow, I have not been able to get it through conformity assessment. Which options do I have to offer to my customer? Please consider all possible options.

The name “OpenAI o1” reflects a new direction in OpenAI’s model development, marking a fresh start from the existing GPT series. The “o1” name indicates the beginning of this new series, with the “o” standing for the focus on reasoning and advanced problem-solving capabilities. This reset to “1” signifies a new line of models, distinct from the GPT-4 series, aimed at tackling complex reasoning tasks through techniques like chain-of-thought reasoning and reinforcement learning. What this means is that once GPT-5 comes out, it will exist side by side with the o1 model.

GPT-3, 4, 5 and so on are the base models. They are trained on a vast amount of data. In ChatMDR, they are augmented and tuned with MDCG and MDR (or IVDR for ChatIVDR) data.

o1 is a new way to use these models. The o1 model is designed to take more time to think through complex tasks. This allows it to perform better on challenging problems in areas like math, coding, and logic-heavy domains​ such as a medical device regulation.

One of the key innovations of the o1 model is its use of “chain-of-thought reasoning.” This technique enables the model to break down a complex problem into smaller, manageable steps before providing a response. This method helps the model refine its thought process and arrive at more accurate solutions, similar to how humans might work through multi-step reasoning problems​.