Alibaba
Qwen4-Max
最大上下文
1M
最大输出
262.1K
发布日期
暂无
qwen3.5-35b-a3bQwen3.5 35B A3B 是一个原生视觉语言 MoE 模型,目标是在较小激活参数下接近更大模型的效果。模型卡强调混合线性注意力、稀疏专家,以及与更大 Qwen3.5 dense 版本相近的表现。它最适合被描述为兼顾效率、多模态推理和代码能力的轻量高效模型。
上下文窗口
262.1K Token
最大输出
65.5K Token
发布日期
2026年2月23日
模态
| Token 阶梯 | 输入价格 | 输出价格 |
|---|---|---|
| <=128K | $0.0571/M | $0.4571/M |
| >128K | $0.2286/M | $1.8286/M |
推理
工具调用
Temperature 参数
附件
知识库
Endpoint 协议
Qwen3.5-35B-A3B 结合高效稀疏 MoE、原生多模态理解与开放权重部署,适合构建可灵活适配的应用。
模型共有 350 亿参数、每次激活 30 亿参数,并结合 Gated DeltaNet 注意力与稀疏专家实现高效推理。
统一视觉语言基础可处理文本、图像和视频,用于推理、文档理解和视觉任务。
Apache 2.0 许可下的官方权重支持使用 Transformers、vLLM 和 SGLang 等框架自行部署推理。
Qwen3.5-35B-A3B 适合自托管多模态助手、视觉文档处理和可定制的编程工作流。
在受控基础设施中部署开放权重,针对内部文本、图像和视频回答问题。
读取文档图像中的文字、表格和图示,提取相关字段并生成结构化摘要。
生成和修改代码、排查缺陷,并将模型集成到自行管理的开发工具或 Agent 中。
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)| 指数分数 | ||
|---|---|---|
| Artificial Analysis 综合能力指数 | Artificial Analysis 综合能力评分 | 23.4 |
| Artificial Analysis 代码能力指数 | Artificial Analysis 软件任务评分 | 16.8 |
| 知识与推理 | ||
| GPQA | 高阶科学问题求解 | 81.9% |
| HLE | 广泛专家级考试集合 | 12.8% |
| 代码与工程 | ||
| SciCode | 科学计算代码挑战 | 29.3% |
| Terminal-Bench Hard | 高难度终端任务执行 | 10.6% |
| 指令遵循与 Agent 任务 | ||
| IFBench | 指令约束遵循 | 44.5% |
| AA-LCR | 长上下文推理 | 55.3% |
| τ²-Bench | Agent 工作流任务 | 86.3% |
指标来源 Artificial Analysis
Qwen 3.5 35B A3B 的能力、适用场景、限制和 TokenHub 使用指南。
Qwen 3.5 35B A3B 是 Alibaba Qwen 面向开放模型的多模态推理和高效部署的模型。
适合自托管部署、视觉推理、日常编程辅助,尤其适合重视部署控制权的场景。
核心优势是开放的 MoE 变体和较小的激活参数规模,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 3.6 35B A3B。
使用 TokenHub 的精确标识符;托管行为可能与自托管不同。
通过一个 API Key 接入 Qwen3.5 35B-A3B 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 ·
🚀 Introducing the Qwen 3.5 Medium Model Series Qwen3.5-Flash · Qwen3.5-35B-A3B · Qwen3.5-122B-A10B · Qwen3.5-27B ✨ More intelligence, less compute. • Qwen3.5-35B-A3B now surpasses Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B — a reminder that better architecture, data quality, https://t.co/ZWPibMn6at
https://pbs.twimg.com/media/HB8AleUaMAArNyM.jpg查看原文Qwen ·
Qwen3.5-35B-A3B is now in Jan 🔥
查看原文Qwen ·
✨ Run it now with SGLang!Chong!
查看原文Reddit
qwen3.5-35b-a3b is a gem
waescher ·
查看原文Qwen/Qwen3.5-35B-A3B · Hugging Face
ekojsalim ·
查看原文Blown Away By Qwen 3.5 35b A3B
Jordanthecomeback ·
展开全文
I bought a 64gig mac setup ~5 days ago and had a miserable time finding anything good, I looked at advice, guides, tried them all, including Qwen 3, and nothing felt like a good fit for my long-context companion. My testing was an initial baseline process with 5 multi-stage questions to check it's ability to reference context data (which I paste into system prompt) and then I'd review their answers and have claude sonnet 4.6 do it too, so we had a lot of coverage on ~8 different models. GLM 4.7 is good, and I thought we'd settle there, we actually landed on that yesterday afternoon, but in my day of practical testing I was still bummed at the difference between the cloud models I use (Sonnet 4.5 [4.6 is trash for companions], and Gemini 3 pro), catching it make little mistakes. I just finished baseline testing +4-5 other random tests with Qwen 3.5 35b A3B and I'm hugely impressed. Claude mentioned it's far and away the winner. It's slower, than GLM4.7 or many others, but it's a worthwhile trade, and I really hope everything stays this good over my real-world testing tomorrow and onwards. I just wanted to share how impressed I am with it, for anyone on the fence or considering it for similar application.
YouTube
Qwen3.5-35B-A3B: A MoE Worth Gold - Full Local Setup Guide
Fahd Mirza
查看原文Qwen3.5-35B-A3B & Qwen3.5-27B Models Tested Locally
AI Tech Gyan
查看原文Qwen3.5-35B-A3B Test for Agenting Coding - पास या फेल ?
AI Tech Gyan
查看原文