What is Qwen3.7 Preview
Qwen3.7 Preview is the next-generation flagship large model preview released by the Alibaba Tongyi Qianwen team, including two versions: Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview. The model achieves significant improvements in agentic coding, world knowledge, and instruction following, helping Alibaba rise to 6th in the text domain and 5th in the vision domain on the LMSYS Chatbot Arena leaderboard. Max focuses on extreme complex reasoning and coding capabilities, while Plus emphasizes a balanced experience of million-level long context and Agentic Coding
Main Features of Qwen3.7 Preview
- Qwen3.7 Max Preview
- Flagship-Level Complex Reasoning: Leading performance on mainstream coding benchmarks such as SWE-bench Pro and Terminal-Bench, supporting highly difficult software engineering tasks and multi-step logical reasoning.
- World Knowledge and Instruction Following: Significantly enhanced world knowledge coverage, able to accurately understand and execute complex instructions, reducing the risk of knowledge hallucination and erroneous execution.
- Native Multimodal Understanding: Supports text, image, and video input, with visual reasoning capabilities ranking among the global top five, enabling cross-modal information fusion.
- Ultra-Long Context Processing: Supports a 256K Token context window, capable of codebase-level analysis and deep understanding of long documents without the need for segmented input.
- Hybrid Reasoning Mode: Supports seamless switching between thinking mode and non-thinking mode, flexibly matching different task complexities while balancing depth and efficiency.
- Qwen3.7 Plus Preview
- Million-Token Native Context: Can process an entire code repository or ultra-long documents in one go, enabling end-to-end long-text reasoning and information extraction.
- Agentic Coding Autonomous Programming: Autonomously plans, executes, and optimizes development tasks in complex engineering environments, supporting multi-turn interactive code generation and debugging.
- Balanced Performance and Cost: Achieves comprehensive performance close to the Max version at lower inference cost, suitable for high-frequency calls and enterprise-level production deployment scenarios.
- Multimodal Native Reasoning: Supports mixed input of text, images, and video, enabling cross-modal information fusion and structured output to meet content creation needs.
- Enterprise-Level Agentic AI Optimization: Supports large-scale production environment deployment, deeply integrated with the Alibaba Cloud Bailian ecosystem, providing a stable and reliable commercial access experience.
Technical Principles of Qwen3.7 Preview
- MoE Mixture-of-Experts Architecture: Based on a mixture-of-experts model architecture, achieving high-density model performance with fewer activated parameters, obtaining stronger reasoning capabilities under equivalent computing power.
- Large-Scale Reinforcement Learning Optimization: During the training phase, automatically expanding test cases improves code execution success rates, strengthening the model's reliability in programming and complex tasks.
- Long-Horizon RL (Long-Horizon RL): Encourages the model to solve complex tasks through multi-turn interaction, supports continuous learning and policy optimization, and enhances the Agent's autonomous decision-making capabilities.
- Thinking Budget Control Mechanism: Supports dynamic adjustment of reasoning depth; users can configure the thinking Token budget according to task requirements, balancing response quality and speed.
- Preserve Thinking (Preserve Thinking): Preserve the complete reasoning process in Agent tasks, ensuring continuity and traceability across multi-turn interactions for easier debugging and auditing.
How to use Qwen3.7 Preview
- Visit the Arena evaluation platform: Visit the Arena official website https://arena.ai/，Qwen3.7 Preview has landed on this platform for public head-to-head evaluation.
- Select the model version In the Arena model list, choose Qwen3.7-Max-Preview (ultimate reasoning) or Qwen3.7-Plus-Preview (balanced experience).
- Start conversation testing: Enter text questions to verify language understanding and instruction-following capabilities, or test visual capabilities through multimodal tasks.
- Head-to-head comparative evaluation: Compare side by side with other top models on the same platform.
Core advantages of Qwen3.7 Preview
-
Dual-version parallel strategy: Max focuses on ultimate complex reasoning and programming, while Plus focuses on million-token long context and Agentic Coding, covering tiered needs.
-
Global leaderboard ranking: On LMSYS Chatbot Arena, it helped Alibaba jump to 6th in text and 5th in vision, with programming and multimodal strength gaining international recognition.
-
Leading in programming benchmarks: It performs excellently in mainstream programming evaluations such as SWE-bench Pro and Terminal-Bench, supporting highly difficult software engineering tasks.
-
Ultra-long context window: Max supports 256K tokens, while Plus natively supports million-level tokens, enabling end-to-end processing of code repositories and ultra-long documents.
-
Hybrid reasoning architecture: Seamlessly switch between thinking mode and non-thinking mode within a single model, flexibly matching different task complexities and cost requirements.
-
Native multimodal understanding: Supports mixed input of text, images, and video, with visual reasoning capability among the world's top tier.
Comparison of Qwen3.7 Preview with similar competitors
| Comparison dimension | Qwen3.7 Preview | GPT-5.5 | DeepSeek V4 |
|---|---|---|---|
| Model positioning | Dual-version parallel strategy (Max/Plus), covering ultimate performance and cost-effectiveness | Strongest overall capability, leading in real-time retrieval and tool calling | The cost-effective choice for long context, open source and locally deployable |
| Arena ranking | 6th in text, 5th in vision (Alibaba Lab) | Leading at the top in text/vision | Did not enter the top five |
| Programming capability | Leading domestically on benchmarks such as SWE-bench Pro | Strong overall programming capability | Excellent code generation and mathematical reasoning |
| Context length | Max 256K / Plus 1M Token | Standard context | 1M tokens (expandable) |
| Reasoning mode | Seamless switching between thinking/non-thinking | Supports deep reasoning | Supports thinking mode |
| Real-time retrieval | Relies on external tools | Native real-time web search, high accuracy | Requires external retrieval tools for assistance |
| Pricing strategy | Plus starts at ¥2/million tokens, Max uses tiered pricing | $5-30/million tokens, relatively high cost | Open source free/low-cost API |
| Deployment method | Alibaba Cloud Model Studio/Qwen Studio | OpenAI API/ChatGPT | Open-source weights/local deployment/API |
| Multimodal support | Native text/image/video | Full modality support | Primarily text, partially multimodal |
Application scenarios of Qwen3.7 Preview
-
Intelligent software development: Based on leading programming capabilities on SWE-bench Pro, it assists with code generation, debugging, and repository-level project analysis, supporting highly difficult software engineering tasks.
-
Code repository-level analysis: With the Plus million-token or Max 256K context window, it understands the entire codebase structure in one pass, enabling end-to-end architecture review and optimization suggestions.
-
Enterprise knowledge management: Handles deep understanding and key information extraction from ultra-long contracts, research reports, and technical documents, maintaining global logical coherence without the need for segmented input.
-
Multimodal content analysis: Integrates text, image, and video inputs for cross-modal reasoning, suitable for visual content understanding, video summary generation, and multimedia material review.
-
Automated agent construction: Relying on Agentic Coding and hybrid reasoning modes, it builds complex business automation workflows capable of autonomous planning, multi-turn interaction, and invoking external tools.