來源Qwen (GitHub Releases: qwen-code)•2026年5月13日較早收集於 31h

Qwen-Code v0.15.11-preview.2：CLI 與效能優化更新

閱讀原文： Qwen (GitHub Releases: qwen-code)

#cli-tools#developer-experience#structured-outputqwen-codeqwenanthropicripgrep

透過全新的結構化 CLI 輸出與自動化 PR 風險分析工具，提升您的開發效率。

30 秒速覽

有什麼變化

新增 --json-schema 支援，實現無頭模式下的結構化輸出。

為什麼重要

這些更新透過在 CLI 環境中直接啟用結構化數據提取與自動化代碼審查功能，簡化了開發人員的工作流程。改進的快取與效能優化降低了高頻 LLM 互動的延遲。

下一步行動

在您的無頭 CLI 工作流程中測試新的 --json-schema 旗標，以標準化自動化管線的 LLM 輸出解析。

誰應關注：Developers & AI Engineers

關鍵要點

•新增 --json-schema 支援，實現無頭模式下的結構化輸出。
•引入 codegraph 技能，用於自動化 PR 審查風險分析與衝突檢測。
•提升 Anthropic 代理相容性，並啟用全域提示詞快取範圍。
•優化核心效能，限制會話列表元數據讀取並實作延遲訊息計數。

深度解析

背景與延伸：來自公開資料，非原文內容。引用 15 個來源。

增強重點摘要

•The --json-schema support in headless mode is specifically designed for automation and scripting workflows, including CI/CD pipelines, allowing Qwen Code to return structured output (text or JSON) programmatically for seamless integration.
•Qwen-Code functions as an open-source AI agent for the terminal, initially forked from Google's Gemini CLI, and is specifically optimized for the Qwen series models, particularly Qwen3-Coder.
•The newly introduced codegraph skill significantly enhances agentic programming workflows by enabling automated Pull Request (PR) handling, rebases, and formatting, contributing to deep code understanding and editing across large codebases.
•Qwen-Code offers multi-protocol flexibility, allowing developers to integrate with various API providers including OpenAI, Anthropic, and Gemini-compatible APIs, as well as Alibaba Cloud's native Qwen endpoints, thereby reducing vendor lock-in.
•Core performance optimizations in this release, such as bounding session-list metadata reads to the head/tail 64KB and implementing lazy message counting, aim to improve efficiency and responsiveness, especially for multi-step automation and long-running tasks.

競品分析

Type

Qwen-Code: Open-source CLI agent
GitHub Copilot: Proprietary IDE extension
Gemini Code Assist: Proprietary CLI/IDE agent
Mistral Vibe: Proprietary agentic platform
DeepSeek-Coder-V2: Open-source LLM

Primary Use

Qwen-Code: Agentic coding, PR analysis, codebase understanding, automation
GitHub Copilot: Code completions, generation
Gemini Code Assist: Code generation, debugging, context-aware assistance
Mistral Vibe: Code generation, testing, deployment, debugging, documentation
DeepSeek-Coder-V2: Code generation, mathematical reasoning

Integration

Qwen-Code: Terminal-first, IDE (VS Code, Zed, JetBrains), SDK
GitHub Copilot: IDE (VS Code, JetBrains)
Gemini Code Assist: IDE (VS Code, JetBrains), Gemini CLI
Mistral Vibe: Terminal, IDE extensions, automated agents
DeepSeek-Coder-V2: Model (integrated into other tools)

Context Window

Qwen-Code: 256K native, up to 1M with extrapolation (Qwen3-Coder)
GitHub Copilot: Not explicitly detailed, but generally large
Gemini Code Assist: Not explicitly detailed, but designed for full codebase context
Mistral Vibe: Understands full codebase context
DeepSeek-Coder-V2: Not explicitly detailed, but trained on 6 trillion tokens

Pricing/License

Qwen-Code: Apache 2.0 for framework/models; API key/Coding Plan for usage
GitHub Copilot: Subscription-based
Gemini Code Assist: Free tier (6,000 daily requests), usage-based for professional
Mistral Vibe: Not explicitly detailed, but platform-based
DeepSeek-Coder-V2: Open-source

Key Differentiator

Qwen-Code: Open-source agentic CLI, multi-protocol API support, deep code understanding, PR risk analysis
GitHub Copilot: Real-time code completions and suggestions
Gemini Code Assist: Generous free tier, deep codebase context, Google ecosystem integration
Mistral Vibe: Specialized AI coding models, multi-agent architecture (iFlow, similar concept)
DeepSeek-Coder-V2: MoE architecture (236B total params, 21B active), extensive training data, multi-language support

Feature / Product	Qwen-Code	GitHub Copilot	Gemini Code Assist	Mistral Vibe	DeepSeek-Coder-V2
Type	Open-source CLI agent	Proprietary IDE extension	Proprietary CLI/IDE agent	Proprietary agentic platform	Open-source LLM
Primary Use	Agentic coding, PR analysis, codebase understanding, automation	Code completions, generation	Code generation, debugging, context-aware assistance	Code generation, testing, deployment, debugging, documentation	Code generation, mathematical reasoning
Integration	Terminal-first, IDE (VS Code, Zed, JetBrains), SDK	IDE (VS Code, JetBrains)	IDE (VS Code, JetBrains), Gemini CLI	Terminal, IDE extensions, automated agents	Model (integrated into other tools)
Context Window	256K native, up to 1M with extrapolation (Qwen3-Coder)	Not explicitly detailed, but generally large	Not explicitly detailed, but designed for full codebase context	Understands full codebase context	Not explicitly detailed, but trained on 6 trillion tokens
Pricing/License	Apache 2.0 for framework/models; API key/Coding Plan for usage	Subscription-based	Free tier (6,000 daily requests), usage-based for professional	Not explicitly detailed, but platform-based	Open-source
Key Differentiator	Open-source agentic CLI, multi-protocol API support, deep code understanding, PR risk analysis	Real-time code completions and suggestions	Generous free tier, deep codebase context, Google ecosystem integration	Specialized AI coding models, multi-agent architecture (iFlow, similar concept)	MoE architecture (236B total params, 21B active), extensive training data, multi-language support

技術深入

Qwen-Code is a command-line interface (CLI) tool optimized for the Qwen series of large language models, particularly Qwen3-Coder.
The underlying Qwen models are built on a transformer-based architecture, incorporating innovations in attention mechanisms and training methodologies.
Qwen3-Coder, the primary model leveraged by Qwen-Code, utilizes a Mixture-of-Experts (MoE) architecture, with its flagship variant boasting 480 billion total parameters and 35 billion active parameters.
It supports a native context window of 256,000 tokens, which can be extended up to 1 million tokens using extrapolation methods, enabling it to handle repository-scale operations and dynamic data like pull requests.
The interaction flow within Qwen-Code involves user input via the CLI, which is then processed by a core package responsible for constructing prompts, interacting with the configured model API, and managing tool execution.
Tool execution, especially for operations that modify the file system or execute shell commands, typically requires explicit user approval for security.
Configuration settings for Qwen-Code are managed through JSON files, environment variables, and command-line arguments, allowing for flexible customization across different scopes (user, project, system).

前景展望基於引用來源的 AI 分析

Enhanced automation of developer workflows will accelerate software delivery.

The new codegraph skill for PR risk analysis and structured output in headless mode directly supports more sophisticated CI/CD pipelines and scripting, enabling greater automation in development and deployment processes.

Increased adoption in diverse enterprise environments is likely.

Structured output capabilities and improved proxy compatibility for Anthropic models make Qwen-Code more suitable for integration into varied enterprise tech stacks and automated systems, broadening its appeal beyond individual developers.

Developers will experience greater flexibility and reduced vendor lock-in in their choice of AI coding models.

The multi-protocol support for various APIs (OpenAI, Anthropic, Gemini, Alibaba Cloud) allows developers to seamlessly leverage different models based on specific task requirements, performance needs, or cost considerations.

時間線

2023-04

Alibaba launched a beta of Qwen (Tongyi Qianwen).

2023-09

Qwen opened for public use after regulatory clearance.

2025-07

Qwen3-Coder, an open-source agentic code model and the foundation for Qwen-Code, was released.

2025-12

Qwen Code v0.5.0 released, introducing Visual Studio Code integration.

2026-04-01

Qwen3.6-Plus launched, featuring a 1M context window and significantly improved agentic coding capabilities.

2026-04-15

Qwen OAuth free tier was discontinued, requiring users to switch to API Key or Coding Plan for authentication.

2023-04
Alibaba launched a beta of Qwen (Tongyi Qianwen).
2023-09
Qwen opened for public use after regulatory clearance.
2025-07
Qwen3-Coder, an open-source agentic code model and the foundation for Qwen-Code, was released.
2025-12
Qwen Code v0.5.0 released, introducing Visual Studio Code integration.
2026-04-01
Qwen3.6-Plus launched, featuring a 1M context window and significantly improved agentic coding capabilities.
2026-04-15
Qwen OAuth free tier was discontinued, requiring users to switch to API Key or Coding Plan for authentication.