Alibaba's Qwen3.7-Max is here — a new flagship agent model that just pulled off something pretty wild. Without any chip architecture docs or performance data, it spent 35 hours autonomously optimizing kernel code, making 1,158 tool calls along the way. The result? A 10x speedup for Triton operators running on Alibaba's in-house Pingtouge Zhenwu M890 processor.
The optimization journey unfolded in five key stages. First, it used Split-K partitioning to split the prefix KV-cache along the token dimension, filling all 36 SM cores. Then it swapped synchronous cudaMalloc for pre-allocated PyTorch variables and eliminated synchronous cudaMemcpy for querying prefix length by using tensor metadata, completely removing host-device communication overhead. Finally, it restructured the operator to handle all four query tokens in a single thread block, sharing loads to amortize memory access overhead — a critical architectural specialization.
Benchmark results show Qwen3.7-Max achieved a geometric mean speedup of 10.0x, significantly outpacing GLM 5.1 (7.3x) and Kimi K2.6 (5.0x). DeepSeek V4 Pro managed just 3.3x and prematurely ended the task after five consecutive rounds without any tool calls.
To generalize across diverse environments, Qwen3.7-Max decoupled tasks, runtime frameworks, and validators during training, using cross-framework reinforcement learning to avoid shortcut overfitting on specific benchmarks. On general agent benchmarks MCP-Mark (60.8 points) and SpreadSheetBench (87.0 points), Qwen3.7-Max shows strong generalization, with overall performance approaching Claude-4.6-Opus-Max.