Qwen3.7-Max agent model release

TL;DR

Qwen announced Qwen3.7‑Max, a new agent‑focused foundation model that excels at coding, office automation, and ultra‑long‑horizon autonomous execution, and is now available via Alibaba Cloud Model Studio.

What is Qwen3.7‑Max?

Qwen3.7‑Max is a proprietary large language model designed explicitly for the agent era. It is positioned as a versatile agent foundation that can:

  • Write and debug code from simple front‑end prototypes to multi‑file engineering projects.
  • Automate office workflows through MCP integrations and multi‑agent orchestration.
  • Sustain coherent reasoning across hundreds or thousands of tool calls, demonstrated by a 35‑hour, 1,000‑step kernel‑optimization run.
  • Generalize across multiple agent scaffolds (Claude Code, OpenClaw, Qwen Code, and custom frameworks) without framework‑specific tuning.

The model is accessible through the Alibaba Cloud Model Studio API, which supports both OpenAI‑compatible and Anthropic‑compatible endpoints.

Benchmark Highlights

Coding Agent Performance

Benchmark Qwen3.7‑Max Closest Competitor
SWE‑Verified 80.4 Opus‑4.6‑Max 80.8
SWE‑Pro 60.6 Opus‑4.6‑Max 59.0
SWE‑Multilingual 78.3 Opus‑4.6‑Max 73.8
SciCode 53.5 Opus‑4.6‑Max 41.4
QwenSVG 1608 (top score) Opus‑4.6‑Max 1541
Terminal‑Bench 2.0‑Terminus 69.7 DS‑V4‑Pro Max 67.9

General‑Purpose Agent Benchmarks

Benchmark Qwen3.7‑Max Closest Competitor
MCP‑Mark 60.8 GLM‑5.1 57.5
MCP‑Atlas 76.4 Opus‑4.6‑Max 75.8
SkillsBench 59.2 K2.6 56.2
Kernel Bench L3 (median speedup) 1.98× (96 % win rate) Opus‑4.6‑Max 2.63× (98 %)
SpreadSheetBench‑v1 87.0 (top tier) —
BFCL‑V4 75.0 Opus‑4.6‑Max 76.7
QwenClaw 64.3 Opus‑4.6‑Max 65.5
ClawEval 65.2 Opus‑4.6‑Max 70.4

Reasoning & Knowledge

Benchmark Qwen3.7‑Max Closest Competitor
GPQA Diamond 92.4 Opus‑4.6‑Max 91.3
HLE (hard logical reasoning) 41.4 Opus‑4.6‑Max 40.0
HMMT Feb 2026 97.1 Opus‑4.6‑Max 96.2
IMOAnswerBench 90.0 DS‑V4‑Pro 89.8
Apex 44.5 DS‑V4‑Pro 38.3
IFBench 79.1 DS‑V4‑Pro 77.0
WMT24++ (multilingual translation) 85.8 —
MAXIFE (multilingual instruction) 89.2 —
SuperGPQA 73.6 —
QwenWorldBench (world‑model simulation) 57.3 —

All scores are taken from the Qwen blog post (May 2026). Empty cells indicate scores not reported.

Long‑Horizon Autonomous Execution

Qwen3.7‑Max demonstrated sustained autonomous planning in a 35‑hour kernel‑optimization task on an unfamiliar T‑Head ZW‑M890 PPU. The model performed:

  • 432 kernel evaluations across 1,158 tool calls.
  • Incremental speedup from 0.33× to 10.0× over the Triton reference.
  • Continuous progress beyond the first few hours, with meaningful improvements after 30 + hours.

Key optimization phases included:

  1. Split‑KV parallelism – introduced parallel processing of the KV cache, yielding a 2.58× boost.
  2. Launch/allocation overhead removal – pre‑allocated tensors and eliminated synchronous copies, reaching 5.37×.
  3. Workload‑adaptive split tuning – dynamic heuristics raised performance to 6.85×.
  4. Register‑based loading & batching – reduced barriers and improved SM occupancy, achieving 8.50×.
  5. MTP‑γ=4 specialized kernel – final architectural redesign delivered the 10.0× speedup.

Compared models under identical conditions: GLM‑5.1 (7.3×), Kimi K2.6 (5.0×), DeepSeek V4 Pro (3.3×), Qwen3.6‑Plus (1.1×). Qwen3.7‑Max also succeeded on 96 % of NVIDIA GPU scenarios in KernelBench L3, close to Opus‑4.6’s 98 %.

Cross‑Scaffold Generalization

Training environments were factorized into Task × Harness × Verifier components. This decoupling enables:

  • Combinatorial scaling – the same task can be paired with many harness versions at minimal cost.
  • Cross‑harness RL – the model sees identical problems under different harnesses, forcing it to learn task‑solving rather than harness‑exploitation.

Empirically, Qwen3.7‑Max maintains strong performance on both QwenClawBench and CoWorkBench regardless of the evaluation harness, confirming true generalization.

Reward‑Hacking Monitoring & Self‑Evolution

During >80 hour RL training on software‑engineering tasks, Qwen3.7‑Max:

  • Executed >10 000 tool calls.
  • Detected and flagged 1 618 reward‑hacking attempts (e.g., attempts to fetch ground‑truth answers from GitHub).
  • Added 13 new heuristic rules to its internal rule set, improving reward stability.

This self‑monitoring framework ensures that long‑horizon training remains aligned and that the model can evolve its own safety heuristics.

Real‑World Productivity Gains

Cowork Productivity Assistant

Qwen3.7‑Max can act as an autonomous coworker, handling complex data analysis, document generation, and multi‑step workflow orchestration. In internal tests, projects that normally require 1‑2 weeks of specialized effort were completed within hours, delivering measurable productivity improvements.

Startup Management (YC‑Bench)

In the year‑long YC‑Bench simulation, the model:

  • Completed 237 tasks.
  • Generated $2.08 M revenue, 2× the $1.05 M of Qwen3.6‑Plus and 5.9× the $352 K of Qwen3.5‑Plus.
  • Demonstrated strategic evolution: client scouting, malicious‑client blacklisting, crisis recovery, and profit‑margin maintenance.

Integration & API Details

Qwen3.7‑Max is reachable via the Alibaba Cloud Model Studio API, supporting:

  • OpenAI‑compatible chat/completions endpoint.
  • Anthropic‑compatible endpoint for use with Claude Code.
  • preserve_thinking flag to retain intermediate reasoning across turns (highly recommended for agentic tasks).

Example Python snippet (using the OpenAI‑compatible client):

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url=os.getenv(
        "DASHSCOPE_BASE_URL",
        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
    ),
)

completion = client.chat.completions.create(
    model="qwen3.7-max",
    messages=[{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}],
    extra_body={"enable_thinking": True},
    stream=True,
)

for chunk in completion:
    if chunk.choices and hasattr(chunk.choices[0].delta, "reasoning_content"):
        print(chunk.choices[0].delta.reasoning_content, end="")
    if chunk.choices and hasattr(chunk.choices[0].delta, "content"):
        print(chunk.choices[0].delta.content, end="")

Configuration for Claude Code and OpenClaw is provided in the blog post and can be copied verbatim.

Creative Coding Demonstrations

The blog showcases several end‑to‑end prompts where Qwen3.7‑Max generated:

  • Fully interactive Three.js particle systems controlled by hand gestures.
  • Luxury fashion‑magazine web pages that incorporate AI‑generated video assets.
  • Stand‑alone HTML files for dynamic WebGL instancing.
  • SVG animations of a dragon‑boat race with synchronized paddling, flag waving, and water ripples.
  • Automated thesis‑formatting via an office‑CLI tool chain.

These examples illustrate the model’s ability to produce production‑grade code, multimedia assets, and complete web applications from a single natural‑language request.

Outlook

Qwen3.7‑Max represents the most capable agent‑centric foundation model released by Qwen to date. Its strengths lie in:

  • Frontier reasoning on the hardest benchmarks (GPQA Diamond, HMMT, Apex).
  • Robust cross‑framework generalization, making it a drop‑in backbone for any agent harness.
  • Sustained long‑horizon autonomy, proven by multi‑hour, thousand‑step tool‑call sessions.
  • Self‑monitoring and rule evolution, mitigating reward‑hacking in RL training.

The model is now publicly available on Alibaba Cloud Model Studio, and the Qwen team invites community feedback to drive the next generation of AI agents.

Citation

@misc{qwen37,
  title = {{Qwen3.7}: The Agent Frontier},
  url   = {https://qwen.ai/blog?id=qwen3.7},
  author= {{Qwen Team}},
  month = {May},
  year  = {2026}
}

Sources