QN 3.6 Plus - Agentic Coding Model¶

Summary¶

QN（通义千问）发布了新旗舰模型 QN 3.6 Plus（2026年4月20日上线），专注于 agentic coding（智能体编码）、实用推理和完成真实工作，而非仅仅"听起来聪明"。该模型在代码编写、仓库级任务、前端生成方面表现优秀，侧重于输出实用性而非空话。

Key Points¶

1. 100万 token 上下文窗口：默认支持 100 万 token，在大型项目中保持稳定性和连贯性
2. 免费使用：通过 QN Code + Q Oath 可获得每天 1000 次免费请求
3. QN Code 开源：编码代理工具本身是开源的，降低使用门槛
4. OpenRouter 免费路由：也可通过 OpenRouter 免费试用，无需更换工作流
5. 非开放权重：旗舰 Plus 版本为闭源，但阿里巴巴表示会继续支持开源社区的 QN 3.6 模型

Chapters¶

00:00 Introduction — QN 3.6 Plus 发布介绍
00:30 模型定位 — Agentic coding、实用推理、真实工作
00:57 上下文窗口 — 100 万 token，大型项目中的优势
01:28 免费使用 — QN Code + Q Oath 每天 1000 次免费请求
02:45 QN Code 介绍 — 开源终端编码代理
03:15 开源 vs 闭源 — 工具层开源，模型权重闭源
03:34 OpenRouter 可用 — 免费路由，无需更换工作流
04:53 AI 编码工具定价 — 当前市场乱象分析
05:22 Agentic Coding 趋势 — 超越一次性代码补全
05:40 开放权重讨论 — Plus 闭源，但社区仍将获支持
06:22 实用性评估 — 闭源但高度可访问
07:07 阿里巴巴开源承诺 — 选中尺寸的 QN 3.6 模型持续开源
07:35 总结推荐 — 模型强、免费、易用
08:09 结尾 — 评论与互动

Full Transcript (with timestamps)¶

0:02 [music]
0:05 Hi, welcome to another video. So, Qen
0:08 has launched another new model and this
0:11 one is QN 3.6 Plus. And to be honest,
0:15 this looks like a really good model for
0:17 sure, especially if you care about
0:19 coding, repo level tasks, front-end
0:21 generation, and generally getting useful
0:24 output instead of fluff. This was made
0:27 live on April 20, 2026.
0:30 And from what QN is saying officially,
0:32 this model is focused very heavily on
0:34 agentic coding, practical reasoning, and
0:37 completing real work rather than just
0:40 sounding smart. And that matters a lot
0:42 because many models can explain things
0:44 nicely. But when it comes to actually
0:46 navigating a code base, planning steps,
0:49 fixing issues, or staying coherent
0:51 across a longer task, that is where they
0:54 start struggling. Now, one of the first
0:56 things that stands out here is the
0:57 context window. QN 3.6 Plus comes with a
1:01 1 million token context window by
1:03 default, which is kind of crazy. And I
1:05 know that context numbers are often used
1:07 as marketing, but if the model can
1:09 actually stay useful with long context,
1:11 then that becomes a huge advantage
1:13 because once your project gets bigger,
1:16 once there are multiple files, existing
1:18 patterns, tests, docs, and edge cases, a
1:21 lot of models start losing the plot. So
1:23 if this one can stay stable over longer
1:25 sessions, that is a really big deal.
1:28 But honestly, the biggest reason I like
1:30 this launch is not just that the model
1:32 seems strong. It is that people can
1:34 actually use it very easily. Because
1:37 many times a company launches a good
1:39 model and then the access is weird, the
1:41 pricing is bad, the provider support is
1:44 limited, or you have to jump through a
1:46 bunch of annoying steps just to see if
1:48 it is any good. So even if the model is
1:50 technically impressive, a lot of normal
1:53 users never really try it. Here, that is
1:55 not really the case. If you use QN code,
1:58 you can use QN 3.6 Plus for fully free
2:01 through Q Oath, which is super nice. And
2:04 this is not just some tiny teaser
2:06 either. Right now, QN code officially
2:09 says that the OOTH flow gives you 1,000
2:11 free requests per day. So for students,
2:14 indie devs, hobbyists, or even
2:16 professionals who just want to test the
2:17 model properly before paying for
2:19 anything, this is really great. And I
2:22 think that is a very smart move. If your
2:24 model is actually good, then let people
2:27 use it. Let them test it on real repos.
2:29 Let them compare it against Claude,
2:31 Gemini, GPT, Deepseek or whatever else
2:33 they already use. That is how trust gets
2:37 built. Dia, that is how people figure
2:39 out whether a model is actually useful
2:41 for their workflow instead of just
2:43 looking good in benchmarks. Now, if you
2:45 do not know what Qen code is, it is
2:48 basically QN's open- source coding agent
2:51 that lives in your terminal. type /
2:54 select Q oath sign in with your QN
2:57 account and then you're ready to go. So
2:59 there is not a ton of friction there and
3:01 that matters a lot more than some people
3:03 think because if setup is annoying most
3:06 people keep saying they will try it
3:07 later and then they never do but if the
3:09 path is basically install sign in and
3:11 start using it then many more people
3:13 will actually give it a serious shot and
3:15 the nice thing here is that QN code
3:17 itself is open source now I will come
3:20 back to the fact that QN 3.6 6 plus is
3:22 not open weights because that is
3:24 definitely important but at least the
3:25 tool layer around it is open source and
3:27 visible. So there is still something
3:30 quite developer friendly about the
3:31 overall package. Now it is not just QN
3:34 code either. QN 3.6 plus is also
3:37 available for free on open router which
3:39 is awesome as well. So if you already
3:40 use open router with your editor coding
3:42 assistant or some custom setup then you
3:45 can try this model there too. And that
3:46 is a big deal because it means you do
3:48 not have to completely change your
3:49 workflow just to evaluate one model.
3:52 That is one of the biggest barriers with
3:54 new models. Sometimes the model may be
3:56 strong, but if you need to move
3:57 everything around, switch tools, rebuild
4:00 configs, and change your setup just to
4:02 try it once, then most people will skip
4:04 it. But if it shows up on Open Router
4:07 and there is a free route there, then
4:09 suddenly it becomes much easier to test.
4:11 You can plug it into the tools you
4:13 already use and find out very quickly
4:15 whether it is actually good for your
4:16 tasks and that is exactly what I like
4:18 here. Qen 3.6 plus is not being
4:21 presented like some inaccessible
4:23 enterprise model that only works inside
4:25 one locked ecosystem. You can use it
4:27 inside Q code for free and you can also
4:29 try it on open router for free. That
4:31 makes it far easier for people to judge
4:32 the model based on results rather than
4:34 just hype. Of course, with free routes
4:36 on Open Router, there can be limits,
4:38 temporary availability changes, and rate
4:40 constraints. So, keep that mind. Free
4:43 does not always mean unlimited, and it
4:45 does not always mean perfectly stable
4:47 access all the time. But still, the fact
4:49 that a model like this is available
4:51 there for free is genuinely quite great.
4:53 And this is where the value conversation
4:55 becomes really important because if
4:57 you're into AI coding tools right now,
4:59 the pricing situation is honestly all
5:01 over the place. Some tools are very good
5:03 but expensive. Some are cheaper but not
5:05 that great. Some give you a strong model
5:08 but wrap it in a workflow you do not
5:10 like. And some look nice in demos but
5:12 then just burn credits when you use them
5:14 for actual work. So whenever something
5:16 strong comes along and people can try it
5:18 seriously without paying first that
5:20 automatically makes it more interesting.
5:22 Now another reason I think this model
5:24 matters is because coding models are
5:26 becoming more and more agentic. It is no
5:29 longer just about oneshot code
5:30 completions. So generally the better
5:33 models are expected to understand the
5:34 the understand the goal as well as
5:36 inspect. Qen as a brand has built a
5:40 strong reputation around open models
5:42 too. A lot of people associate QN with
5:45 useful open releases that can be run in
5:47 more flexible ways. So when they put out
5:49 a flagship model like this and it is not
5:52 open weights that definitely stands out.
5:54 It is not shocking because many
5:56 companies keep their best flagship
5:58 models closed, but it is still notable
6:00 and worth pointing out clearly. If you
6:03 are someone who only wants fully
6:04 openweight models, if you care deeply
6:06 about self-hosting, local deployment,
6:09 full control, or running the exact same
6:11 model on your own infrastructure, then
6:13 this is obviously a downside. There is
6:16 no point pretending otherwise. If open
6:18 weights are a requirement, then Q13.6 6
6:21 plus being closed is simply a
6:23 limitation. But if your main question is
6:25 can I use a really good coding model
6:27 right now without paying immediately
6:29 then this still looks like a very
6:30 compelling option because even though
6:33 the flagship model itself is not open
6:34 weights access to it is still
6:37 surprisingly open in practice. You are
6:39 not being forced into some absurd
6:41 pricing wall just to see whether it is
6:42 good. You can actually try it and that
6:45 changes the conversation a lot. That is
6:48 why I said this is interesting. It is
6:50 closed but it is still highly
6:51 accessible. QN code itself is open
6:54 source. The model is available through
6:56 QN OOTH for free inside QN code. It is
7:00 also available as a free route on open
7:02 router. So from a user access
7:04 perspective, this is much better than
7:06 what you often get with closed flagship
7:07 models. Also, Alibaba has said they will
7:10 continue to support the open source
7:12 community with selected QN 3.6 models in
7:15 developerfriendly sizes. So it is not
7:18 like they are abandoning open models
7:20 entirely. It is more that the plus
7:22 flagship here is closed while the
7:24 broader QN ecosystem still keeps some
7:26 open source direction. Install the tool,
7:28 sign in and start using it or open
7:31 router and test the free route in your
7:33 existing setup. That is a pretty
7:35 compelling story. So overall I think Qen
7:38 is doing a lot right here. The model
7:40 looks strong. The focus on coding and
7:42 reasoning is exactly where demand is. So
7:45 the free access and QN code is very
7:47 useful. The open router availability is
7:49 also great and even though it is not
7:51 open weights, it is still much easier to
7:53 try than many other flagship level
7:54 models. That combination is what makes
7:57 this feel easy to recommend. It is not
7:59 open weights which is definitely
8:01 interesting and definitely worth
8:02 pointing out. But as a practical model
8:04 to actually use, this seems quite
8:06 awesome. Overall, it's pretty cool.
8:09 Anyway, let me know your thoughts in the
8:11 comments. If you like this video,
8:12 consider donating through the super
8:14 thanks option or becoming a member by
8:16 clicking the join button. Also, give
8:18 this [music] video a thumbs up and
8:19 subscribe to my channel. I'll see you in
8:21 the next one. Until then, bye.