GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

Qwen3.7-Max Leads SWE-Pro at 60.6% — Alibaba's Coding Agent Runs 35 Hours, 1,000+ Tool Calls

Alibaba released Qwen3.7-Max on May 19, positioning it explicitly as a model for the “agent era” rather than a general-purpose assistant. The headline benchmark is SWE-bench Pro at 60.6%, which leads every model in Qwen’s comparison table — Kimi K2.6 at 59.5%, DeepSeek V4 Pro at 59.0%, GLM-5.1 at 58.8%, and Claude Opus 4.6 at 57.3%.

Key Numbers

BenchmarkQwen3.7-MaxKimi K2.6DS-V4-ProOpus 4.6
SWE-bench Pro60.6%59.5%59.0%57.3%
SWE-bench Verified80.4%80.2%80.6%80.8%
SWE-bench Multilingual78.3%76.7%76.2%77.5%
Terminal-Bench 2.0 (Terminus)69.7%66.7%67.9%65.4%
GPQA Diamond92.4%90.5%90.1%91.3%
HLE41.4%36.4%37.7%40.0%

The Terminal-Bench numbers use a different harness (5-hour timeout, Terminus-2) than the main leaderboard rankings, so they are not directly comparable to the scores from vix (90.2%) or JJAgent (87.1%), which use the standard scaffold.

The Agent Design

Qwen3.7-Max is built around sustained execution. Alibaba’s most cited demonstration is a 35-hour continuous kernel optimisation run that made over 1,000 tool calls without human intervention. Kernel Bench L3 shows a 1.98x median speedup and 96% of problems faster than torch.compile — second only to Claude Opus 4.6 (2.63x / 98%) in their benchmark set, but competitive.

The model is designed to generalise across agent scaffolds. Alibaba specifically tested it across Claude Code, OpenClaw, and Qwen Code, citing consistent performance rather than scaffold-specific tuning. MCP integrations and multi-agent orchestration are listed as primary deployment targets, with CoWorkBench (67.2%) and MCP-Atlas (76.4%) showing strength in workflow automation.

General Reasoning

On STEM benchmarks, Qwen3.7-Max hits 41.4% on HLE (Human Last Exam) with tools, narrowly ahead of Kimi K2.6 (40.0%) and comfortably ahead of DeepSeek V4 Pro (38.3%). On HMMT 2026 February mathematics competition, it reaches 97.1%, the highest in the comparison table.

GPQA Diamond at 92.4% also leads the table — the only model to clear 92.

What Is Not Known

Pricing is not announced. The model will be available via Alibaba Cloud Model Studio but as of launch is listed as “coming soon.” SWE-bench Verified (80.4%) lands below DeepSeek V4 Pro (80.6%) and Claude Opus 4.6 (80.8%), meaning on the main leaderboard SWE metric, it sits in the cluster rather than leading it. The SWE-Pro edge is meaningful but comes from a benchmark Qwen runs internally.

Arena ELO is not yet available — the leaderboard added “qwen3.7-max-preview” on May 14 but vote counts will take weeks to stabilise.