阿里千問推出Qwen Coding Plan
💡Qwen Coding Plan adds top models & tool compatibility—supercharge your AI coding workflow.
⚡ 30-Second TL;DR
有什麼變化
推出Qwen Coding Plan以強化模型支援
為什麼重要
此計劃擴展Qwen在編碼任務的生態系，讓開發者更容易整合熱門工具。它強化阿里在AI編碼領域的競爭地位。
下一步行動
Test Qwen3-Coder-Next integration with ClaudeCode via the new Qwen Coding Plan.
關鍵要點
- •推出Qwen Coding Plan以強化模型支援
- •上新Qwen3.5-Plus及Qwen3-Coder-Next模型
- •相容QwenCode、ClaudeCode、Cline等編碼工具
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 8 個來源。
🔑 增強重點摘要
- •Alibaba's Qwen Coding Plan is a subscription service from Alibaba Cloud Model Studio offering up to 90,000 requests per month for AI coding tools like Qwen Code, Claude Code, Cline, and OpenClaw[2].
- •The plan supports new models including qwen3.5-plus, qwen3-max-2026-01-23, qwen3-coder-next, qwen3-coder-plus, and third-party models like glm-4.7 and kimi-k2.5[2].
- •Qwen3-Coder-Next is an open-weight model with 80B total parameters (3B activated via MoE), 256K context length, designed for local coding agents on consumer hardware[1][3][5].
- •Qwen3.5-Plus features hybrid reasoning (Thinking/Fast/Auto modes), up to 1M token context, 480B MoE architecture with 35B active parameters, and multimodal capabilities[1][4][7].
- •Qwen coding models excel in benchmarks: Qwen2.5-Coder at 88.4% HumanEval (beats GPT-4's 87.1%), 69.6% SWE-Bench, supporting 119 languages and local deployment on 32GB RAM[1].
📊 競品分析▸ Show
| Feature | Qwen Coding Plan / Qwen3-Coder-Next | Competitors (e.g., GPT-4o, Claude Sonnet 4.5) |
|---|---|---|
| Benchmarks | HumanEval 88.4%, SWE-Bench 69.6%; Sonnet 4.5-level[1][3] | GPT-4: HumanEval 87.1%; Claude competitive[1] |
| Parameters | 80B total, 3B active (MoE); 480B MoE (35B active)[1][3] | Proprietary, larger active params typically[1] |
| Context Length | 256K-1M tokens[1][3][4] | Varies, e.g., Claude up to 200K[3] |
| Pricing | Fixed monthly fee, 90K requests; free open-source local[1][2] | Per-token API; subscription tiers[1] |
| Deployment | Local on 32-64GB RAM/consumer GPU; open-weight[1][3] | Cloud-only proprietary[1] |
🛠️ 技術深入
- Qwen3-Coder-Next: 80B total parameters, 3B activated per inference using Mixture-of-Experts (MoE) with 512 experts (10 activated per token), Hybrid architecture (Gated DeltaNet + MoE + Gated Attention), 256K context, trained on large-scale executable tasks + RL, open-weight Apache-2.0 license[3][5][6].
- Qwen3.5-Plus: 480B MoE with 35B active parameters, hybrid reasoning (Thinking mode with 1024 token budget, Fast, Auto with tools), 1M token context, 250K vocabulary, multi-token prediction (10-60% token savings), native multimodal (text/images/UI), 119 languages including programming[1][4][7].
- Coding Plan Models: Includes qwen3-max-2026-01-23 (thinking enabled), supports third-party integration via Model Context Protocol (MCP) for agents[2].
🔮 前景展望AI analysis grounded in cited sources
Qwen Coding Plan democratizes advanced AI coding with affordable subscriptions, open-weight local models, and superior benchmarks, challenging proprietary leaders like OpenAI and Anthropic by enabling privacy-focused, cost-effective agentic coding on consumer hardware, potentially accelerating open-source adoption in enterprise development[1][2][3].
⏳ 時間線
📎 來源 (8)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 36氪 ↗
每週 AI 簡報
每週一封，可隨時退訂。