Alibaba
Qwen4-Max
最大上下文
1M
最大输出
262.1K
发布日期
暂无
qwen3.6-plusQwen3.6 Plus 的介绍重点在混合线性注意力与稀疏 MoE 架构,并强调 Agentic Coding、前端开发和推理任务的提升。公开模型卡提到仓库级任务、交互式前端生成,以及 SWE 类评测表现。它的定位更接近开发者与生产力模型,而不只是普通对话助手。
上下文窗口
1M Token
最大输出
65.5K Token
发布日期
2026年4月2日
模态
| Token 阶梯 | 输入价格 | 输出价格 | 缓存创建 5m | 缓存读取 5m |
|---|---|---|---|---|
| <=256K | $0.2857/M | $1.7143/M | $0.3571/M | $0.0286/M |
| >256K | $1.1429/M | $6.8571/M | $1.4286/M | $0.1143/M |
推理
工具调用
Temperature 参数
附件
知识库
Endpoint 协议
Qwen3.6-Plus 结合智能体编程、原生视觉理解与百万级上下文窗口,适合处理复杂开发任务。
增强的代码能力可用于仓库级工程、前端开发和多轮迭代式编程。
可处理文本、图像和视频,并增强了物体识别、OCR 与目标定位能力。
百万级上下文窗口可容纳大型代码库、长篇文档和较长的多模态输入。
Qwen3.6-Plus 适合仓库级软件开发、基于视觉参考的前端实现和长篇多模态资料分析。
分析大型代码库中的依赖关系,规划协同修改,并生成跨多个文件的实现。
理解截图或设计参考中的布局和视觉细节,并将其转换为前端代码。
审阅包含文本、图像和视频的长篇资料,识别物体、提取可见文字并汇总相关结论。
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)Qwen3.6 Plus
| 指数分数 | ||
|---|---|---|
| Artificial Analysis 综合能力指数 | Artificial Analysis 综合能力评分 | 39.6 |
| Artificial Analysis 代码能力指数 | Artificial Analysis 软件任务评分 | 42.9 |
| 知识与推理 | ||
| GPQA | 高阶科学问题求解 | 88.2% |
| HLE | 广泛专家级考试集合 | 25.7% |
| 代码与工程 | ||
| SciCode | 科学计算代码挑战 | 40.7% |
| Terminal-Bench Hard | 高难度终端任务执行 | 43.9% |
| 指令遵循与 Agent 任务 | ||
| IFBench | 指令约束遵循 | 75.2% |
| AA-LCR | 长上下文推理 | 69.7% |
| τ²-Bench | Agent 工作流任务 | 97.7% |
指标来源 Artificial Analysis
Qwen 3.6 Plus 的能力、适用场景、限制和 TokenHub 使用指南。
Qwen 3.6 Plus 是 Alibaba Qwen 面向均衡的多模态推理和生产任务的模型。
适合图像和视频理解、文档分析、大量工具调用的自动化,尤其适合重视可靠的多步骤执行的场景。
核心优势是质量、多模态能力和生产功能之间的良好平衡,并具备可在深度思考和直接回答之间切换的混合思考。
它属于较早一代,可能缺少后续模型的新能力。 更看重需要最新能力时,可考虑 Qwen 3.7 Plus。
使用 TokenHub 显示的精确标识符,并按账户文档确认当前功能。
通过一个 API Key 接入 Qwen3.6 Plus 与更多 AI 模型。
媒体与讨论
精选与此模型相关的公开视频和帖子。
X (Twitter)
Qwen ·
(1/8)🚀 Introducing Qwen3.6-Plus: Towards Real-World Agents! 🤖 Today, we’re thrilled to drop a major milestone in our journey toward native multimodal agents. Here is what makes Qwen3.6-Plus a game-changer: 💻 Next-level Agentic Coding: Smarter, faster execution. 👁️ https://t.co/uUpcyLaB6d
https://pbs.twimg.com/media/HE5885PbgAAlo4y.jpg查看原文Qwen ·
🚀 Introducing Qwen3.6-Max-Preview, an early preview of our next flagship model Highlights: ⚡️ Improved agentic coding capability over Qwen3.6-Plus 📖 Stronger world knowledge and instruction following 🌍 Improved real-world agent and knowledge reliability performance Smarter, https://t.co/0Fr8jgqDbJ
https://pbs.twimg.com/media/HGWpSykaMAAqsJn.jpg查看原文Together AI ·
Introducing Qwen3.6-Plus from @Alibaba_Qwen, a 1M-context model built for real-world agents, agentic coding, and multimodal reasoning. AI natives can now use Qwen3.6-Plus on Together AI and benefit from reliable inference for production-scale agent workflows. https://t.co/s4Ysm1ZsKt
https://pbs.twimg.com/media/HHF9k19bQAA8aaM.jpg查看原文Reddit
[deleted by user]
查看原文Alibaba's Qwen3.6-Plus is beating Claude Opus in coding!!
AdVirtual2648 ·
展开全文
alibaba just dropped qwen 3.6-plus and the benchmarks are kind of ridiculous. it's scoring 61.6 on terminal-bench and 57.1 on swe-bench verified. for context that puts it ahead of claude 4.5 opus, kimi k2.5, and gemini 3 pro on most of the agentic coding tests. the crazy part is it's less than half the size of kimi k2.5 and glm-5. way smaller model but matching or beating the big ones. it also has a native 1M context window which is huge if you're working on long codebases or big document tasks. and they built it specifically for agentic workflows so it's not just "generate code and hope for the best"... it actually handles multi-step tasks. it's already free on openrouter too. open source versions coming soon apparently. link's in the comments.
qwen/qwen3.6-plus is testing very well in my setup
productboy ·
展开全文
I'm about ten days into hardcore testing of Hermes with different setups. What I've tried: Hetzner VPS with 4vCPU, 8GB memory. Free models via OpenRouter. It was very unreliable because of rate limiting of free models. Hermes would start strong on a task then stall with 429 errors. Hetzner VPS with 4vCPU, 8GB memory. Free models via Ollama running on the VPS. This was unusable because of the small models that had to be used given VPS memory constraints. Hermes needs larger models to perform reliably. Oracle VPS with 1vCPU, 1GB memory. Had to try it but Hermes wouldn't run and barely installed. Hetnzer VPS with 4vCPU, 8GB memory. Paid models via OpenRouter. The default model is qwen3.6-plus with a gemma model as fallback. This has been the best performing setup. And, it really shows off the self-improvement architecture where Hermes seems to get better with each session. The Discord integration is also more stable with higher fidelity output.
YouTube
Qwen3.6-Plus : New AI Model is Breaking the Internet!
Codedigipt
查看原文Qwen3.6-Plus - The #1 AI Model?
Blunt AI - Technology through Data
查看原文Qwen3.6-Plus: Advancing Real-World Agentic Intelligence
LuxaK
查看原文