After a night of tinkering with the local big model, we finally chose to release it: 1. Main model: Qwen-30B-Instruction, it is sufficient for daily work. (Instructions follow very strictly) 2. Inference backup: An 8-bit modified GPT OSS mlx is kept. The 4-bit version of GPT-OSS is not very effective, and it is not meaningful to fully allocate three levels of inference budget. 3. Coder for all directions, planning to use SOTA flagship directly without considering local models (after all, it's just for work)