The recent introduction of models like Qwen 8B DeepSeek R1 has sparked great interest in the artificial intelligence community, thanks to its remarkable reasoning capabilities, which exceed expectations considering the number of parameters.

The problem of diffusion

The question that arises spontaneously is: why don't we see a greater diffusion of "distilled" models of this type? These models, capable of functioning effectively even on less performing hardware, offer a significant advantage in terms of accessibility and costs.

For those evaluating on-premise deployments, there are trade-offs between performance and costs that AI-RADAR helps evaluate through its analytical frameworks on /llm-onpremise.

General context

The "distillation" of complex models is a technique that allows the transfer of knowledge from a large model (the "teacher") to a smaller one (the "student"). This process makes it possible to obtain more efficient models suitable for scenarios with limited computational resources, paving the way for new applications in the edge and on-premise fields.