Recently, Mistral, a domestic large-scale model released by Tsinghua University, has aroused enthusiastic responses on GitHub. Its 2B size has performance that surpasses many "large-scale" models, which is amazing. This is not only reflected in its powerful performance, but also in its extremely low cost advantage: the inference cost of 1,700,000 tokens can be obtained with only 1 yuan, which is far lower than similar products. In addition, Mistral also has multi-modal capabilities, showing strong application potential. This incident once again proves that in the field of AI, excellent model design and cost control are equally crucial, and it is not simply "volume is king."
The article focuses on:
Recently, the Tsinghua University Department released a domestically produced Mistral. This large model, which is only 2B in size, unexpectedly received a warm welcome on GitHub and gained 300+ stars in one day. In terms of performance, this product is quite competitive. There is a great contrast between performance and size. It has surpassed many mainstream "large-volume" large models in many achievements. The cost aspect is even more amazing. It only costs 1 yuan to obtain 1,700,000 tokens at the inference cost. Compared with similar products, the cost is much lower. In addition to the above features, the product also has multi-modal capabilities and exhibits excellent results.
The success of Mistral demonstrates the breakthrough in performance and cost of domestic large models, and also provides new ideas for the development direction of large models in the future. I believe that in the future, we will see more similar surprises appear, promoting the continuous progress of AI technology.