According to CoinMeta, AI Memory Company, in collaboration with legal firm AI and company Harvey, conducted an experiment to create a virtual law firm that includes 100 million token, nearly 10,000 documents, and 266 legal cases. The 27B Qwen model achieved an average score of 67% on 250 legal tasks, surpassing all other models tested. The cost per query is approximately 0.13 US dollars, which is only 1/10 to 1/24 of the 4.8 high-intensity reasoning required by Claude Opus. Before training, the accuracy rate of this model was 4.7%, but it increased to 72.6% after training. The model is capable of more efficient targeted searches during tasks, with 99.2% of cases allowing for direct access to relevant information from the beginning.