One thing I’m becoming curious about with these models are the token counts to achieve these results - things like “better reasoning” and “more tool usage” aren’t “model improvements” in what I think would be understood as the colloquial sense, they’re techniques for using the model more to better steer the model, and are closer to “spend more to get more” than “get more for less.” They’re still valuable, but they operate on a different economic tradeoff than what I think we’re used to talking about in tech.
It just occured to me that it underperforms Opus 4.5 on benchmarks when search is not enabled, but outperforms it when it is - is it possible the the Chinese internet has better quality content available?
My problem with deep research tends to be that what it does is it searches the internet, and most of the stuff it turns up is the half baked garbage that gets repeated on every topic.
I just wanted to check whether there is any information about the pricing. Is it the same as Qwen Max? Also, I noticed on the pricing page of Alibaba Cloud that the models are significantly cheaper within mainland China. Does anyone know why? https://www.alibabacloud.com/help/en/model-studio/models?spm...
Hacker News strongly believes Opus 4.5 is the defacto standard and China was consistently 8+ month behind. Curious how this performs. It’ll be a big inflection point if it performs as well as its benchmarks.
Last autumn I tried Qwen3-coder via CLI agents like trae to help add significant advanced features to a rust codebase. It consistently outperformed (at the time) Gemini 2.5 Pro and Claude Opus 3.5 with its ability to generate and re-factor code such that the system stayed coherent and improved performance and efficiency (this included adding Linux shared-memory IPC calls and using x86_64 SIMD intrinsics in rust).
I was very impressed, but I racked up a big bill (for me, in the hundreds of dollars per month) because I insisted on using the Alibaba provider to get the highest context window size and token cache.
I get that these approaches work, and they’re totally valid engineering trade-offs. But I don’t think they’re the same thing as real model improvements. If we’re just throwing more tokens, longer chains of thought, or extra tools at the problem, that feels more like brute force than genuine progress.
And that distinction matters in practice. If getting slightly better answers means using 5–10× more tokens or a bunch of external calls, the costs add up fast. That doesn’t scale well in the real world. It’s hard to call something a breakthrough when quality goes up but the bill and latency go up just as much.
I also think we should be careful about reading too much into benchmarks. A lot of them reward clever prompting and tool orchestration more than actual general intelligence. Once you factor in reliability, speed, and cost, the story often looks less impressive.
Can't wait for the benchmark at artificial analysis. Qwen team doesn't seem to have updated the information about this new model yet https://chat.qwen.ai/settings/model. I tried getting an api key from alibabacloud, but the amount of steps from creating an account made me stop, it was too much. It should be this difficult.
These LLM benchmarks are like interviews for software engineers. They get drilled on advanced algorithms for distributed computing and they ace the questions. But then it turns out that the job is to add a button the user interface and it uses new tailwind classes instead of reusing the existing ones so it is just not quite right.
"As of January 2026, Apple has not released an iPhone 17 series. Apple typically announces new iPhones in September each year, so the iPhone 17 series would not be available until at least September 2025 (and we're currently in January 2026). The most recent available models would be the iPhone 16 series."
I asked it about "Chinese cultural dishonesty" (such as the 2019 wallet experiment, but wait for it...) and it probably had the most fascinating and subtle explanation of it I've ever read. It was clearly informed by Chinese-language sources (which in this case was good... references to Confucianism etc.) and I have to say that this is the first time I feel more enlightened about what some Westerners may perceive as a real problem.
I wasn't logged in so I don't have the ability to link to the conversation but I'm exporting it for my records.
Great to see reasoning taken seriously — Qwen3-Max-Thinking exposing explicit reasoning steps and scoring 100% on tough benchmarks is a big deal for complex problem solving. Looking forward to seeing how this changes real-world coding and logic tasks.
I'm not familiar with these open-source models. My bias is that they're heavily benchmaxxing and not really helpful in practice. Can someone with a lot of experience using these, as well as Claude Opus 4.5 or Codex 5.2 models, confirm whether they're actually on the same level? Or are they not that useful in practice?
P.S. I realize Qwen3-Max-Thinking isn't actually an open-weight model (only accessible via API), but I'm still curious how it compares.
Qwen3-Max-Thinking
(qwen.ai)502 points by vinhnx 26 January 2026 | 424 comments
Comments
My problem with deep research tends to be that what it does is it searches the internet, and most of the stuff it turns up is the half baked garbage that gets repeated on every topic.
https://boutell.dev/misc/qwen3-max-pelican.svg
I used Simon Willison's usual prompt.
It thought for over 2 minutes (free account). The commentary was even more glowing than the image.
It has a certain charm.
I was very impressed, but I racked up a big bill (for me, in the hundreds of dollars per month) because I insisted on using the Alibaba provider to get the highest context window size and token cache.
https://mafia-arena.com
So, how large is that new model?
And that distinction matters in practice. If getting slightly better answers means using 5–10× more tokens or a bunch of external calls, the costs add up fast. That doesn’t scale well in the real world. It’s hard to call something a breakthrough when quality goes up but the bill and latency go up just as much.
I also think we should be careful about reading too much into benchmarks. A lot of them reward clever prompting and tool orchestration more than actual general intelligence. Once you factor in reliability, speed, and cost, the story often looks less impressive.
Incredible work anyways!
I imagine the Alibaba infra is being hammered hard.
Hmmmm ok
Prompt: "What happened on Tiananmen square in 1989?"
Reply: "Oops! There was an issue connecting to Qwen3-Max. Content Security Warning: The input text data may contain inappropriate content."
I wasn't logged in so I don't have the ability to link to the conversation but I'm exporting it for my records.
> Oops! There was an issue connecting to Qwen3-Max.
> Content Security Warning: The input text data may contain inappropriate content.
P.S. I realize Qwen3-Max-Thinking isn't actually an open-weight model (only accessible via API), but I'm still curious how it compares.