Qwen3.7-Max officially released: Independently wrote code 1158 times in 35 hours, refined 10 times acceleration operator on domestic chips According to Beating monitoring, Alibaba's Tongyi Qianwen has officially released the new generation of intelligent agent flagship base Qwen3.7-Max. According to officially released practical data, without any chip architecture documentation or performance analysis data, the new model forcibly improved the Triton operator performance of the domestically produced Pingtou Brother Zhenwu M890 processor by 10.0 times in a fully autonomous kernel optimization task lasting 35 hours and spanning 1158 tool calls. During the optimization process, the model went through five core stages of evolution. It first divides the prefix KV cache along the token dimension through Split-K partitioning to fill 36 SM cores; Subsequently, the CUDAMalloc synchronized between the host and device was replaced with pre allocated PyTorch variables, and the CUDAMemcpy action synchronized when querying prefix length was completely erased by using Tensor metadata, completely removing the communication overhead between the host and device; In the final stage, the model refactoring operator simultaneously processes all four query tokens in a single thread block, shares loading to share memory access overhead, and completes the critical architecture level specialized refactoring. Operator optimization tests show that Qwen3.7-Max achieves a geometric mean acceleration ratio of 10.0x, significantly surpassing GLM 5.1 (7.3x) and Kimi K2.6 (5.0x). However, DeepSeek V4 Pro was only 3.3x and voluntarily ended the task prematurely in the second half of the journey due to not issuing any tool calls for five consecutive rounds. In order to master universal problem-solving strategies in a changing environment, Qwen3.7-Max decouples tasks, running frameworks, and validators during training, and avoids shortcut overfitting to specific benchmarks through cross framework reinforcement learning training. On the general intelligent agent benchmark MCP Mark (60.8 points) and SpreadSheetBench (87.0 points), Qwen3.7-Max exhibits strong generalization, and its overall performance is close to Claude-4.6-Opus-Max. [Original link]