Qwen3.7-Max Officially Released: 1,158 self-written codes in 35 hours, developing a 10x accelerated operator on domestic chips

According to Beating Monitoring, Alibaba Tongyi Qianwen officially released the new generation intelligent agent flagship base model Qwen3.7-Max. Official practical data shows that, without any chip architecture documentation or performance analysis data, the new model forcibly improved the performance of the Triton operator on the domestically produced Hengxian Ge True Wu M890 processor by 10.0 times in a fully autonomous kernel optimization task lasting 35 hours and involving 1,158 tool calls.

During the optimization process, the model underwent five core evolution stages. It first partitioned the prefix KV-cache along the token dimension using Split-K to fill 36 SM cores; then replaced the cudaMalloc used for synchronization between host and device with pre-allocated PyTorch variables, and completely eliminated the cudaMemcpy synchronization actions during prefix length queries by using tensor metadata, thereby removing communication overhead between host and device; in the final stage, the model reconstructed operators to process all 4 query tokens simultaneously within a single thread block, sharing load to distribute memory access costs, completing a key architecture-level specialized reconstruction.

Operator optimization tests showed that Qwen3.7-Max achieved a 10.0x geometric mean speedup, significantly surpassing GLM 5.1 (7.3x) and Kimi K2.6 (5.0x). Meanwhile, DeepSeek V4 Pro only achieved a 3.3x speedup and proactively ended the task early in the latter half after five consecutive rounds without any tool calls.

To master general problem-solving strategies in variable environments, Qwen3.7-Max decoupled tasks, runtime frameworks, and validators during training, and avoided overfitting to specific benchmarks through cross-framework reinforcement learning training. On general intelligent agent benchmarks MCP-Mark (60.8 points) and SpreadSheetBench (87.0 points), Qwen3.7-Max demonstrated strong generalization ability, with overall performance closely approaching Claude-4.6-Opus-Max.