On April 2, Alibaba released Qwen3.6-Plus.
This is not a routine release of a "bigger model," but a targeted upgrade specifically for enterprise-level AI agent scenarios. From a product positioning perspective, Alibaba has realized one thing: it cannot win the arms race for general-purpose capabilities, so it will instead build an execution layer that can "truly work in production environments."
Three Most Important Changes
1. Code Capability: Repository-Level Engineering
Previous large models writing code were essentially "here's a function, you use it." Qwen3.6-Plus has a different positioning — it aims to work within an entire code repository: planning, writing code, running tests, finding issues, and modifying code — a complete execution loop.
This working mode is relatively close to Claude Code; the model is not a "suggestion provider" in a prompt, but an active participant.
It natively supports a 1 million token context window, handling the complete code of medium-sized projects with little pressure.
2. Visual Understanding: Screenshots Directly to Code
Given a UI screenshot, a hand-drawn wireframe, or a product prototype, it can directly generate runnable front-end code. This is not generating a rough framework, but a closed loop of "reading the image and writing the code."
Beyond static images, it also supports understanding documents and video content, which is quite practical for internal enterprise tool scenarios.
3. Tool Calling and Long-Horizon Planning
This is currently where the biggest gap in AI agent capabilities lies among various companies. Qwen3.6-Plus achieved top results on multiple long-horizon planning benchmarks, and its accuracy in tool calling has also improved.
In terms of specific numbers, it shows competitive performance on programming-specific evaluations like SWE-Bench and Terminal-Bench, but Alibaba has not fully disclosed all benchmark data.
A New API Feature: preserve_thinking
This parameter maintains the reasoning context across multi-turn conversations, rather than re-reasoning each turn.
It is very practical for agent scenarios: reducing redundant reasoning equals reducing token consumption, which equals lower actual usage costs. In long-flow automation tasks, this difference is amplified.
Why Is This Called a "Alibaba Internal-Use First" Model?
Qwen3.6-Plus will be integrated into two internal Alibaba products:
- Wukong: Alibaba's AI-native enterprise platform, using multi-agent systems to automate complex business tasks
- Qwen App: Alibaba's flagship consumer-facing AI application
Using it internally first, then opening it up externally — this logic often means the model has undergone stress tests that are closer to real-world scenarios than purely external benchmarks.
Programming environment compatibility is good: it supports OpenClaw, Claude Code, Qwen Code, and Kilo Code, while also being compatible with both OpenAI and Anthropic API protocols, so developers do not need major changes to integrate.
Horizontal Comparison: Is Qwen on the Right Path?
Looking at the evolution of the Qwen series — from general conversation to long context, then to code specialization and multimodality, and now to agentic execution — each step moves toward being more "capable of doing work."
This direction increasingly resembles the product strategies of Anthropic (Claude Code) and OpenAI (Codex). The difference is that Alibaba has the inherent advantage of Chinese enterprise customers and deep integration with the Alibaba Cloud ecosystem — this is its differentiation space in the domestic market, winning not through benchmark scores but through ecosystem integration.
Sources: Qwen3.6-Plus: Towards Real World Agents (Alibaba Cloud Community); Alibaba Unveils Qwen3.6-Plus to Accelerate Agentic AI Deployment for Enterprises (Alibaba Cloud); CocoLoop; Beyond Passive Assistance: How Qwen3.6-Plus Is Advancing Agentic AI (The Tech Revolutionist)