In short
Mistral has opened access to a model with one trillion parameters and allows it to be fine-tuned. I analyze how such freedom is useful and why claims of competing with the leaders should still be treated cautiously for now.
Free access to a large model may prove more important than a high-profile comparison with the leaders: Mistral Large 4 is permitted to be used and customized. But the claim that it can compete with the best models from the US and China remains, for now, Mistral’s assertion.
The trillion-parameter model is currently available in a preview version. The final release is promised by the end of the month. For those who need to tailor AI to their tasks, the ability to modify the model offers more freedom than access only through a ready-made service.
Mistral is emphasizing programming and cyber defense, as well as tasks for manufacturing, finance, and electrical engineering. The company explains this choice by saying that different industries require specialized models. This is an understandable direction: a general-purpose model may not be the best tool for every specific job.
But the article does not provide test results that would substantiate the comparison with the leaders. The model is still in preview access, and the final version has not yet been released. Based on this material, it is impossible to assess how well it handles the stated tasks or exactly what can be customized.
For which tasks is it more important for you to be able to tailor AI to your needs than to use the strongest model “out of the box”?
Source: AI - Ars Technica field