Alibaba
Qwen4-Max
最大上下文
1M
最大输出
262.1K
发布日期
暂无
qwen3.6-35b-a3bQwen3.6 35B A3B 是一个激活参数较小的 MoE 模型:相关来源描述为 35B 总参数、约 3B 激活参数,并具备可扩展到接近 1M token 的长上下文能力。官方资料还提到它在代码任务上可以与更大的 dense 模型竞争。它的核心价值是部署效率高,同时保留不错的代码和长上下文能力。
上下文窗口
262.1K Token
最大输出
65.5K Token
发布日期
2026年4月17日
模态
| 输入价格 | 输出价格 |
|---|---|
| $0.2571/M | $1.5429/M |
推理
工具调用
Temperature 参数
附件
知识库
Endpoint 协议
Qwen3.6-35B-A3B 结合开放权重的稀疏 MoE 架构、原生视觉语言处理与增强的智能体编程能力。
混合线性注意力与稀疏专家混合架构在约 350 亿参数中激活约 30 亿参数,以提高推理效率。
原生视觉语言架构可接收文本、图像和视频,并生成基于视觉内容的文本结果。
该后训练开放权重模型增强了前端工作流和仓库级推理,并支持独立部署。
Qwen3.6-35B-A3B 适合独立部署的代码智能体、仓库分析以及针对图像或视频的视觉推理任务。
在受控环境中部署开放权重,通过迭代式智能体流程生成、修改和调试代码。
追踪代码仓库中跨文件的关系,解释实现结构,并制定协同修改方案。
分析图像或视频以定位目标、理解空间关系,并输出结构化文本说明。
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 | 33 |
| Artificial Analysis 代码能力指数 | Artificial Analysis 软件任务评分 | 35.2 |
| 知识与推理 | ||
| GPQA | 高阶科学问题求解 | 84.1% |
| HLE | 广泛专家级考试集合 | 20.2% |
| 代码与工程 | ||
| SciCode | 科学计算代码挑战 | 35.8% |
| Terminal-Bench Hard | 高难度终端任务执行 | 34.8% |
| 指令遵循与 Agent 任务 | ||
| IFBench | 指令约束遵循 | 64.4% |
| AA-LCR | 长上下文推理 | 63.7% |
| τ²-Bench | Agent 工作流任务 | 95.3% |
指标来源 Artificial Analysis
Qwen 3.6 35B A3B 的能力、适用场景、限制和 TokenHub 使用指南。
Qwen 3.6 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、图像和视频理解、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。
开放部署需要自行承担基础设施、推理服务和评估工作。 更看重重视稳定的生产表现时,可考虑 Qwen 3.6 Plus。
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 Qwen3.6 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 ·
⚡ Meet Qwen3.6-35B-A3B:Now Open-Source!🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes https://t.co/UMiChPaLid
https://pbs.twimg.com/media/HGB1pOdbIAA_kyz.jpg查看原文Qwen ·
LM Performance:Qwen3.6-35B-A3B outperforms the dense 27B-param Qwen3.5-27B on several key coding benchmarks and dramatically surpasses its direct predecessor Qwen3.5-35B-A3B, especially on agentic coding and reasoning tasks. https://t.co/PyXDNruoy2
https://pbs.twimg.com/media/HGB1yASbcAEzWkd.jpg查看原文Unsloth AI ·
2-bit Qwen3.6-35B-A3B did a complete repo bug hunt with evidence, repro, fixes, tests and a PR writeup. 🔥 Run it locally in Unsloth Studio with just 13GB RAM. 2-bit Qwen3.6 GGUF made 30+ tool calls, searched 20 sites and executed Python code. GitHub: https://github.com/unslothai/unsloth https://t.co/1YwMxrWOYr
https://pbs.twimg.com/amplify_video_thumb/2044849663774425088/img/fNsqbHmhlS3uC_47.jpg查看原文Reddit
Qwen3.6-35b-a3b seems like the best coding model rn
JSVD2 ·
what are your thoughts on this and real life cases and example?
查看原文Qwen 3.6 35b a3b is INSANE even for VRAM-constrained systems
Lucerys1Velaryon ·
展开全文
Apologies for the 1345th post glazing QWEN but it has literally been a game-changer for me. I’m relatively new to running local LLMs, but the appeal of having a private AI assistant for coding and experimentation is significant. As a software engineer, I’ve also embraced the recent “vibe-coding” trend, mostly using AI to build hobby Android apps in my spare time. Over the past month, I created a web scraper app designed to extract images from ad-heavy websites that are otherwise frustrating to navigate. The project was entirely AI-generated. I didn’t write the code myself, though I did design the architecture and suggest optimizations, which AI still struggles with. I’ll admit I haven’t reviewed most of the code; it’s messy, but since it’s purely a personal hobby project, that’s acceptable. Until recently, I relied on services like Antigravity and Codex, which worked well enough on free plans. However, tighter usage quotas pushed me toward local models. My hardware is modest by local LLM standards: an AMD 7700 XT, 32GB DDR4 RAM, and a Ryzen 5 5600. I experimented with Gemma 3, Gemma 4, and Qwen 2.5 Coder (mostly Q3/Q4 quantizations under 20GB due to VRAM constraints), using LM Studio as the backend with various frontends like GitHub Copilot, RooCode, Cline, Android Studio AI Chat, and OpenCode. Unfortunately, none of these models fit my workflow well. They struggled with even minor bug fixes, frequently exhausted context windows, got stuck in reasoning loops, or failed tool calls repeatedly. Then I tried Qwen 3.6 35B-A3B. Initially, I expected little, but I installed the i1-q4_k_s quant, offloaded all 40 layers to GPU, configured 128k context, enabled flash attention, and used Q8_0 KV quantization. For testing, I gave it a practical task: fix the scraper logic for a problematic website. Gemini Flash (in Antigravity) and MinMax (the free version in Opencode) had both failed to solve this issue despite multiple attempts. Using LM Studio as an OpenAI-compatible endpoint with GitHub Copilot in VS Code, I let it run. It took about 25 minutes, but it succeeded in one shot. With a single initial prompt, it analyzed the site’s HTML structure, compared it against my Kotlin scraper code, and resolved three critical bugs without a single failed tool call. That result was impressive enough that I gave it a second challenge: update the project README with real app screenshots by driving an Android emulator, selecting screenshots based on specific criteria. Using RooCode this time, the process took around 45 minutes. I had to teach it some emulator workflow conventions, such as taking screenshots after every action and analyzing them to track app state, but once instructed, it executed flawlessly. For the first time, I feel like I have a local model capable of reliably handling most of my coding tasks, while reserving cloud-based premium models for more demanding work. Qwen has genuinely made local AI coding practical for me.
Qwen3.6 35B-A3B on a Laptop: My Zero to One Moment
rolznz ·
展开全文
Hi everyone, I'm new here - because I only have a laptop and I only just realized local models are actually good enough now. So I'd like to share my experience, in case it helps others, and also to learn from the more experienced people here. This is the first model that works for me on my ASUS Zenbook Pro 14 (RTX 4060 8GB VRAM, 64GB RAM): fast enough: ~27TPS generation speed at 32k context, or ~18TPS at 256k context smart enough: it can read and write files, use skills, execute CLI commands, use git, follow instructions, and act as a useful thinking partner. Why it's important to me For me this is important because it's where I unconsciously decided to draw the line - that I didn't want to share private information or more personal thoughts with cloud models (even TEE ones). I know I can still get hacked and my data leaked, but for me that's different than giving it up from the first prompt. So for the first time, I now have this fully local, second brain. For me, it's a game changer. I still use cloud models for public stuff I'm still using cloud models for public projects, but for brainstorming and simple personal projects, local is now good enough for me. I'm also now looking into a more powerful desktop machine where maybe I can do some more serious coding. I have had a taste and I want more 😄 Now whenever I see Claude's black box "✽ Envisioning… (41s · ↓ 2.9k tokens · thinking some more with high effort)" it's so frustrating. I have no idea if it's going in the right direction. (whether this is an "efficient" way to do things is another story) My issues so far with Qwen3.6 Qwen3.6 35B A3B is not perfect, here are some minor issues I observed, which I can work around: It makes some mistakes, but normally recovers on its own. Very occasionally it does get stuck in a loop. It does need some human monitoring, which is fine for me. It sometimes doesn't read a skill in full or make the best decision even when it can fit it in context. It seems to sometimes be "lazy". It is very non-deterministic. I didn't do any tweaks here though (because normally it ends up with the result I need). I guess some of these could be improved if I used a larger quantization. My setup For inference I use llama.cpp, with unsloth's Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf. For my harness, I use Pi with pi-llama-cpp extension. The harness runs in multipass and connects to the host running llama.cpp. I've also connected it to my phone through an E2EE Matrix chat (a custom one I built off of pi-messenger-bridge) - although it means I have to keep my laptop on all the time, which is annoying. Another reason for buying another machine which I'm more comfortable to run 24/7. llama.cpp flags for 256k context(18tps): ./build/bin/llama-server -m Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf -ngl 24 -np 1 -fa on -ctk q4_0 -ctv q4_0 -c 262144 --host 0.0.0.0 --port 8088 -ncmoe 32 --no-mmap --jinja llama.cpp flags for the 32k context (27tps): ./build/bin/llama-server -m Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf -ngl 99 -np 1 -fa on -ctk q4_0 -ctv q4_0 -c 32000 --host 0.0.0.0 --port 8088 -ncmoe 32 --no-mmap --jinja What was your Zero to One moment?
YouTube
Qwen3.6 35B-A3B Full Test – Is THIS the Best LOCAL Model Yet?
Bijan Bowen
查看原文Qwen3.6-35B-A3B + OpenClaw - Agentic Coding Locally for Free
Fahd Mirza
查看原文Qwen3.6 Local Test | Can it Beat Gemma 4? | Coding, OCR, Image Understanding with llama.cpp | 🔴 Live
Venelin Valkov
查看原文