Qwen3.7: The Agentic Frontier

Qwen3.7-Max is the latest proprietary model built for the agentic era, unifying coding agents, MCP-integrated office automation, multi-agent orchestration, and long-horizon autonomous execution spanning hundreds to thousands of steps into a single model. It's also positioned as a practical coworker model, handling complex information synthesis, data analysis and modeling, and document/visualization generation.

It will soon be available via API through Alibaba Cloud Model Studio, and is said to work consistently across different agent frameworks such as Claude Code, OpenClaw, and Qwen Code.

Coding agent performance is strong across the board.

• It scored Terminal Bench 2.0-Terminus 69.7, SWE-Pro 60.6, SWE-Multilingual 78.3, SciCode 53.5, and QwenSVG 1608.
• SWE-Verified 80.4 was nearly on par with Opus-4.6 Max (80.8) and DS-V4-Pro Max (80.6).
• In frontend and repository-level tasks, it also showed QwenWebDev 1568 and NL2Repo 47.2.

In general agent tasks, it showed strong generalization independent of the harness used.

• It delivered stable performance on Qwenclaw 64.3, CoWorkBench 67.2, and ClawEval 65.2, and remained near the top on MCP-Mark 60.8, MCP-Atlas 76.4, Skillbench 59.2, BFCL-V4 75.0, and SpreadSheetBench-v1 87.0.
• Notably, on Kernel Bench L3, it achieved a 1.98x median speedup with a 96% win rate, demonstrating strong GPU kernel optimization ability.

Reasoning and general-purpose capabilities were also strong. It recorded GPQA Diamond 92.4, HMMT 2026 Feb 97.1, IMOAnswerBench 90.0, and Apex 44.5, and boosted instruction-following, long-context, and multilingual performance across the board with IFBench 79.1, MMLU-Pro 89.6, MRCR-v2 128k 90.4, WMT24++ 85.8, and MAXIFE 89.2.

The training philosophy centers on environment scaling. Continuing the approach from Qwen3.5, it used a Rollout infrastructure that separates Task-Harness-Verifier, combining diverse harnesses and verifiers, and explained that performance rises steadily together even when evaluated purely on out-of-domain environments unseen during training. This result reportedly led to top-tier average rankings close to Claude-4.6-Opus-Max.

The most impressive case is the Extend Attention kernel optimization. Starting with no documentation, profiling data, or example kernels on a T-Head ZW-M890 PPU environment, it performed 1,158 tool calls and 432 kernel evaluations over about 35 hours, achieving a 10.0x geometric-mean speedup over the Triton baseline.

The improvements progressed as 0.33x→2.58x→5.37x→6.85x→8.50x→10.0x. Along the way, it added Split-KV parallelization and a reduction kernel based on online softmax rescaling, removal of cudaMalloc/cudaFree/cudaMemcpy calls and loop unrolling, workload-adaptive split tuning, register-based K/V loading with batched softmax updates, and finally a dedicated MTP γ=4 kernel.

Under the same conditions, other models only reached GLM 5.1 7.3x, Kimi K2.6 5.0x, DeepSeek V4 Pro 3.3x, and Qwen3.6-Plus 1.1x. Some models terminated early after 5 consecutive rounds without action, and the Qwen team emphasized that long-horizon autonomous optimization can indeed be productive.