Qwen3.7-Max Achieves 10x Performance Improvement on Domestic Chip in 35-Hour Optimization Task

According to Beating Monitoring, Alibaba’s Tongyi Qianwen has officially released its new flagship agent foundation model, Qwen3.7-Max. Official real-world data shows that, without access to any chip architecture documentation or performance analysis data, the new model achieved a 10.0x improvement in Triton operator performance on the domestic Pingtouge Zhenwu M890 processor during a fully autonomous kernel optimization task lasting 35 hours and involving 1,158 tool calls. During optimization, the model underwent five core evolutionary stages. First, it partitioned the prefix KV-cache along the token dimension using Split-K to fully utilize all 36 SM cores. Next, it replaced host-device synchronization via cudaMalloc with pre-allocated PyTorch variables and eliminated all synchronous cudaMemcpy calls used to query prefix lengths by leveraging tensor metadata, completely removing host-device communication overhead. In the final stage, the model restructured the operator to process all four query tokens within a single thread block, sharing memory loads to amortize memory access costs, achieving a critical architecture-specific refinement. Real-world operator optimization results show that Qwen3.7-Max achieved a 10.0x geometric mean speedup, significantly outperforming GLM 5.1 (7.3x) and Kimi K2.6 (5.0x). In contrast, DeepSeek V4 Pro achieved only a 3.3x speedup and prematurely terminated the task after five consecutive rounds with no tool calls. To master generalizable problem-solving strategies in dynamic environments, Qwen3.7-Max decoupled tasks, execution frameworks, and validators during training and employed cross-framework reinforcement learning to avoid shortcut overfitting to specific benchmarks. On the general-purpose agent benchmarks MCP-Mark (60.8) and SpreadSheetBench (87.0), Qwen3.7-Max demonstrated exceptional generalization ability, with overall performance closely approaching that of Claude-4.6-Opus-Max.