Aunt AI: 27B Qwen's legal mission transcends Opus

The AI Memory Company Engram, in collaboration with Law AI, Harvey, created a virtual firm of 100 million tokens, nearly 10,000 documents and 266 legal matters. The 27B Qwen model received an average rating of 67 per cent in 250 legal missions, exceeding all reference models. The cost of a single query was approximately $0.13 and was only 1/10 to 1/24 of Claude Opus 4.8's high-intensity reasoning. Before training, Qwen had an accuracy rate of 4.7 per cent, which increased to 72.6 per cent after training. The model is more effective in conducting targeted searches in missions and 99.2 per cent of missions begin with a direct search for relevant information。