Qwen 3.6 Max (FULLY FREE): Qwen JUST ENDED Opus 4.7? This MODEL is ACTUALLY INSANE!

本站添加: 2026-05-05

[00:00:05]Hi, welcome to another video. So, Quen has launched another new flagship preview and this one is Qen 3.6 Max preview. It was announced today and this looks like Qen's new top-end proprietary model. So, if you were already impressed by Qen 3.6 Plus, this is basically them saying, "Okay, now let's push the flagship even further." Now, let me be very clear about one thing first. This is not an openweight release. So, if you only care about fully open models that you can self-host, then this is obviously not that. But if your main goal is to get the strongest QN model for coding agents, tool use, and general knowledge work, then this looks really interesting for sure. Right now, QN says that you can use it in Q Studio. And they also say that API access is coming under the model name QN36 Max POS. Since this is a preview, the model is still actively evolving. So, take all the benchmark claims with a grain of salt.

[00:01:00]But even with that caveat, the jump over QN 3.6 Plus is pretty noticeable. Now, let's talk about the coding agent side first because that is what most of us care about on SkillsBench QN 3.6. Max preview scores 55.6 compared to 45.7 for Q3.6 Plus. That is a very big jump. On terminal bench 2.0, it goes from 61.6 6 to 65.4. On Q and Claw Bench, it goes from 57.2 to 59.0. On Swebench Pro, it moves from 56.6 to 57.3, which is not a massive jump, but it is still an improvement. And on Qen Webbench, it goes from 1495 LO to 1532 LO, which is also nice to see. So, the story here is pretty clear. QN 3.6 Plus was already positioned as a strong real world coding and agent model, and now QN 3.6 6 max preview is supposed to be the sharper version for harder, longer, more agentic workflows. And honestly, that matters a lot because these days it is not enough for a model to just write one nice function and call it a day. The better models need to inspect a repo, understand what the task actually is, follow the right instructions, survive tool calls, recover from errors, and keep going without completely losing the plot. That is what actually matters in real workflows. And QN seems to be targeting exactly that. Now, what is also interesting is that this is not just a coding update. Quinn is clearly trying to say that Max is also smarter in general and more reliable. On Super GPQA, which is more about graduate level knowledge, it goes from 71.6 on QN 3.6 plus to 73.9. On Q Chinese bench, it jumps from 78.7 to 84.0, which is honestly a very solid increase. And on tool call format ifbench, it moves from 83.3 to 86.1. So instruction following and tool call formatting are also getting better.

[00:02:51]>> They also show AA omniscience index going from 3.0 to 10.0 and GDP vala going from 43.0 to 51.0.

[00:03:01]So their pitch is not just that they made it stronger at coding. Their pitch is more that they made it stronger at knowing things, following instructions properly, and staying reliable in real agent workflows. And that is a much better story than just chasing one benchmark and hoping people get impressed. Now with all of that said, I do not think you should look at this and assume it is just crushing absolutely everything. Some of the charts are clearly more mixed. Claude 4.5 Opus is still ahead on things like Scode and NL2 repo in the image they shared and GLM 5.1 is still very strong on some web and coding benchmarks as well. So this is not one of those launches where I would say okay everyone else is finished. It is more that QN is getting very serious at the high end and the gap between them and the best closed models is now getting small enough that you really have to pay attention and that is what makes this launch actually interesting to me. Quen has already been doing a lot lately. They launched Qen 3.6 plus on April 1st 2026. They followed that up with the open source QN 3.6 6 35B A3B model for people who want something lighter and more deployable. And now on April 20, 2026, they are pushing the flagship preview higher as well. So they are kind of covering the whole stack right now. You have the openweight route if you want flexibility and you have the closed flagship route if you want the strongest performance. I actually like that strategy a lot because one of the annoying things in AI right now is when a company forces you into only one lane.

[00:04:29]either everything is closed and expensive or everything is open, but the flagship edge is just not there. Quen is at least trying to play both sides and for users that is pretty good. Now, the downside is still obvious. This max model is closed. So, again, if open weights are a requirement for you, this is not the release to get excited about.

[00:04:48]Also, it is a preview model, which means the benchmarks can move, the API details can change, availability can change, and the final polished release may still shift. So, I would not build my whole long-term stack around a preview before seeing stable docs, stable pricing, and stable access. But for actual testing and practical usage, this still looks very worth paying attention to. If you already liked Quen 3.6 Plus, but wanted something a bit sharper for harder coding agent tasks, this is probably the model to watch. If you care about tool use, instruction following, and real task completion instead of just pretty answers, this also looks great. And if you have been ignoring Quen because you thought they were mostly benchmark merchants, then I think that is becoming harder and harder to say with a straight face. The biggest takeaway for me is that this does not feel like just another random max label. The improvements are actually pointed in the right direction. More agentic coding, better world knowledge, better instruction following, and better reliability. That is exactly where the frontier models need to improve if they want to be useful beyond demos. So yeah, I think Kuwan is cooked again. Just keep your expectations balanced. This is a preview model. The benchmark image is still mostly internal evaluation data.

[00:05:59]Some competitors are still ahead on a few tasks and we still need to see how this thing feels in long real world sessions, not just in charts. But as a direction, this is really strong. And as a signal that Qen wants to fight seriously at the flagship level, this is pretty hard to ignore. If you can access it in Q Studio, I would absolutely recommend trying it on your own workflow. Throw repo tasks at it. Throw tool use at it. Throw your annoying debugging sessions at it. That is where you will figure out whether it is actually worth the hype for you. For now, I think QN 3.6 Max preview looks like one of the most interesting model launches of April 2026, especially if you care about coding agents and practical AI work instead of just benchmark theater. Overall, it's pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through the super thanks option or becoming a member by clicking the join button. Also, give this video a thumbs up and subscribe to my channel. I'll see you in the next one. Until then, bye.