Qwen3.7-Plus release notes / what's new

Qwen3.7-Plus: A Unified Multimodal Agent Foundation

Qwen3.7-Plus is a multimodal agent model that unifies vision and language into a single foundation, enabling it to operate as a multimodal interactive hybrid agent. It is designed to perceive real-world scenes, operate GUIs, write code from visual references, and navigate mobile applications end-to-end, seamlessly blending GUI and CLI interactions within a single agent loop.

Multimodal Interactive Hybrid Agent Capabilities

Qwen3.7-Plus integrates the workflow of "see, think, write, act, and verify" to automate complex software tasks.

End-to-End Software Development

Using a Hybrid-Agent system based on Qwen3.7-Plus, the team demonstrated the full automation of an English vocabulary learning APP's R&D cycle. The agent operated for over 11 hours, generating more than 10,000 lines of code and triggering over 1,000 agent calls. This process covered requirement document generation, automated coding, deployment, test case creation, GUI-based automated testing, and autonomous version evolution.

Professional Application Replication

The model can autonomously replicate professional desktop applications. In one instance, it recreated a high-fidelity version of the native macOS Stocks app by interacting with the original app to understand UI layout, generating SwiftUI source code, integrating the LongBridge market API for live data, and passing 10 autonomous functional verification tests.

Browser Automation

Through the Qwen for Chrome extension, Qwen3.7-Plus can act as a Browser Agent. It perceives webpages and performs clicks, typing, and navigation to complete complex workflows, such as purchasing the cheapest ECS server via a cloud console, including dynamic adjustments for price changes or inventory limits.

Technical Performance and Benchmarks

Text and General Agent Performance

Qwen3.7-Plus delivers competitive text performance approaching Max-tier models. Key benchmark results include:

  • Coding Agents: Strong performance on Terminal Bench 2.0 (70.3), SWE-bench series, and SciCode (51.3).
  • General Agents: Robust tool-use and planning on MCP-Mark (58.7) and Deep-Planning (62.3), with particular strength in GPU kernel optimization on Kernel Bench L3 (2.06 speedup / 98% faster than torch.compile).
  • STEM & Reasoning: High performance on GPQA Diamond (90.3), HMMT (92.9), and IMOAnswerBench (86.0).

Multimodal Reasoning and Understanding

Qwen3.7-Plus shows systematic enhancements in visual agent capabilities:

  • Multimodal Reasoning: Significant gains on BabyVision (70.4/64.7), MathVision (90.3), and HiPhO (84.1), demonstrating an ability to integrate spatial relationships and physical commonsense.
  • Visual Agent & Coding: Substantial improvements on ScreenSpot Pro (79.0), OSWorld-Verified (73.3), and AndroidWorld (81.0). On QwenVision2Code, it achieves a score of 1772.0, reflecting its ability to turn design references into executable code.
  • Multimodal Search & Knowledge QA: Improvements on SimpleVQA (81.7) and WorldVQA (61.1) by combining visual inputs with external search augmentation.
  • General Visual Understanding: Strong performance in document parsing (OmniDocBench 1.5: 91.4) and OCR (OCR-Bench-V2 ZH: 67.1).

Video and Driving-Scene Understanding

The model reasons over temporal dynamics and semantic relationships in videos, scoring strongly on VideoMMMU (85.4) and TVBench (78.2). In autonomous driving scenarios, it demonstrates spatial relationship understanding on LingoQA (83.4) and SURDS (77.2).

Visual Agent Tool Integration

Qwen3.7-Plus combines visual understanding with programmatic problem solving:

  • Code Interpreter Integration: The model can analyze images to solve sliding-block puzzles, navigate mazes, and assemble jigsaw puzzles by autonomously generating and executing code.
  • Vision-to-Code Generation: It transforms images, videos, and UI screenshots into executable code, including SVG reconstruction and full interactive webpage generation from visual references.

Deployment and Framework Compatibility

Qwen3.7-Plus is available via Alibaba Cloud Model Studio and supports industry-standard protocols compatible with OpenAI's specification. It includes a preserve_thinking feature to maintain thinking content across turns, which is recommended for agentic tasks.

The model generalizes across diverse agent scaffolds and is compatible with:

  • Claude Code: Supports the Anthropic API protocol.
  • OpenClaw: Connects via Model Studio.
  • Qwen Code: Deeply optimized for the Qwen series.

Sources