Qwen3.7 Preview - A preview version of the next-generation flagship large model launched by Ali Tongyi. - AiBoss

What is Qwen3.7 Preview?

Qwen3.7 Preview is a preview version of the next-generation flagship large model launched by Alibaba's Tongyi Qianwen team, including two versions: Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview. The model shows significant improvements in agent programming, world knowledge, and instruction compliance, helping Alibaba jump to 6th place in the text domain and 5th place in the vision domain on the LMSYS Chatbot Arena leaderboard. Max focuses on extremely complex reasoning and programming capabilities, while Plus emphasizes a balanced experience of handling millions of contexts and agentic coding.

Main functions of Qwen3.7 Preview

Qwen3.7 Max Preview
- Flagship-level complex reasoningIt performs well on mainstream programming benchmarks such as SWE-bench Pro and Terminal-Bench, and supports challenging software engineering tasks and multi-step logical reasoning.
- World knowledge and instructionsThe coverage of global knowledge has been significantly enhanced, enabling accurate understanding and execution of complex instructions, reducing the risk of knowledge illusion and mis-execution.
- Native multimodal understandingIt supports text, image, and video input, and its visual reasoning capabilities rank among the top five globally, enabling cross-modal information fusion.
- Long context processingSupports 256K Token context windows, enabling codebase-level analysis and in-depth understanding of long documents without the need for segmented input.
- Hybrid reasoning modeIt supports seamless switching between thinking and non-thinking modes, flexibly matching different task complexities and balancing depth and efficiency.
Qwen3.7 Plus Preview
- Millions of tokens native contextIt can process an entire code repository or an extremely long document at once, enabling end-to-end long text reasoning and information extraction.
- Agentic Coding (Autonomous Programming)It can autonomously plan, execute, and optimize development tasks in complex engineering environments, and supports multi-round interactive code generation and debugging.
- Balancing effectiveness and costLower inference costs achieve near-Max version overall performance, suitable for high-frequency calls and enterprise-level production deployment scenarios.
- Multimodal native inferenceIt supports mixed input of text, images, and video, enabling cross-modal information fusion and structured output to meet content creation needs.
- Enterprise-grade Agentic AI optimizationIt supports large-scale production environment deployment, is deeply integrated with Alibaba Cloud's ecosystem, and provides a stable and reliable commercial access experience.

The technical principles of Qwen3.7 Preview

MoE Hybrid Expert ArchitectureBased on a hybrid expert model architecture, it achieves high-density model performance with fewer activation parameters, and obtains stronger inference capabilities with the same computing power.
Large-scale reinforcement learning optimizationDuring the training phase, the success rate of code execution is improved by automatically expanding test cases, thereby enhancing the reliability of the model in programming and complex tasks.
Long-Horizon Reinforcement LearningIt encourages models to solve complex tasks through multiple rounds of interaction, supports continuous learning and policy optimization, and enhances the agent's autonomous decision-making ability.
Consider budget control mechanismsIt supports dynamically adjusting the inference depth, allowing users to configure the thinking token budget according to task requirements, balancing response quality and speed.
Preserve ThinkingPreserve the complete reasoning process in the Agent task to ensure the continuity and traceability of multi-round interactions, facilitating debugging and auditing.

How to use Qwen 3.7 Preview

Visit the Arena review platformVisit the Arena website https://arena.ai/, where Qwen3.7 Preview is available for public comparison testing.
Select model version In the Arena model list, select either Qwen3.7-Max-Preview (Extreme Reasoning) or Qwen3.7-Plus-Preview (Balanced Experience).
Start the conversation testInput text questions can be used to verify language comprehension and instruction compliance, or multimodal tasks can be used to test visual abilities.
Horizontal comparative evaluation: Compare with other top-level models on the same platform.

Qwen3.7 Preview's core advantages

Dual-version parallel strategyMax focuses on extremely complex reasoning and programming, while Plus focuses on long contexts with millions of tokens and Agentic Coding, covering layered needs.
Global Rankings: Helped Alibaba rise to 6th in text and 5th in visual in LMSYS Chatbot Arena, gaining international recognition for its programming and multimodal capabilities.
Leading Programming BenchmarksIt performs excellently in mainstream programming benchmarks such as SWE-bench Pro and Terminal-Bench, and supports challenging software engineering tasks.
Extra Long Context WindowMax supports 256K tokens, while Plus natively supports millions of tokens, enabling end-to-end processing of code repositories and extremely long documents.
Hybrid Inference ArchitectureSeamlessly switch between thinking and non-thinking modes within a single model to flexibly match different task complexities and cost requirements.
Native multimodal understandingIt supports mixed input of text, images, and videos, and its visual reasoning capabilities rank among the world's top tier.

Comparison of similar products with Qwen3.7 Preview

Comparison Dimensions	Qwen3.7 Preview	GPT-5.5	DeepSeek V4
Model localization	Dual versions running concurrently (Max/Plus), covering both extreme performance and cost-effectiveness.	With the strongest overall capabilities, it leads in real-time retrieval and tool usage.	A cost-effective choice for long-context environments; open-source and locally deployable.
Arena Rankings	Text ranked 6th, visual ranked 5th (Alibaba Labs)	Text/Visual Head Leading	Not in the top five
Programming skills	Domestic Leading Standards in Benchmarks such as SWE-bench Pro	Strong comprehensive programming skills	Excellent code generation and mathematical reasoning
Context length	Max 256K / Plus 1M Token	Standard Context	1M Token (Scalable)
Reasoning patterns	Seamless switching between thinking and non-thinking	Supports deep reasoning	Supports thinking mode
Real-time search	Depends on external tools	Native real-time web search with high accuracy	External search tools are needed.
Pricing Strategy	Plus starting at ¥2 per million tokens, Max tiered pricing.	$5-30 per million tokens, relatively high cost.	Open source and free / low-cost API
Deployment method	Alibaba Cloud Hundred Refinements/Qwen Studio	OpenAI API/ChatGPT	Open source weight/local deployment/API
Multimodal support	Native text/images/videos	Full modal support	Text-based, with some multimodal elements

Application Scenarios of Qwen3.7 Preview

Intelligent software developmentBased on SWE-bench Pro's leading programming capabilities, it assists in code generation, debugging, and repository-level project analysis, supporting highly complex software engineering tasks.
Code repository level analysisUse the Plus Million Tokens or Max 256K context window to understand the entire codebase structure at once and provide end-to-end architecture analysis and optimization suggestions.
Enterprise knowledge managementIt can deeply understand and extract key information from extremely long contracts, research reports, and technical documents, maintaining overall logical coherence without requiring segmented input.
Multimodal content analysisIt integrates text, image, and video inputs for cross-modal reasoning, and is suitable for visual content understanding, video summarization generation, and multimedia material review.
Automated intelligent agent constructionBased on Agentic Coding and hybrid inference, we can build complex business automation processes that can be autonomously planned, interact in multiple rounds, and call external tools.