Leading the pack in both text and vision among domestic models
Hengyu, reporting from Aofeisi
Quantum Bits | WeChat Official Account QbitAI
Alibaba, you really know how to keep quiet — how did Qwen 3.7 Preview suddenly go live in the blink of an eye!
The official Arena account (formerly the LLM leaderboard arena) has posted the freshly released results for Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview.
In text, Qwen3.7-Max-Preview ranks #13, which also lifts Alibaba’s (lab) ranking to #6.
It is also the top-ranked domestic model.
In vision, Qwen3.7-Plus-Preview ranks #16, pushing Alibaba up to #5.
It is likewise the top-ranked domestic model on the leaderboard.
The official Qwen account said it outright: “We can’t wait to release the Qwen3.7 series!”
But wait — wasn’t Qwen3.6-Max-Preview released just at the end of April?
Qwen3.6-Max is still in preview, and Qwen3.7-Max preview is already on the table… It’s hard to tell whether the official release cadence is just unusually slow, or the version iterations are happening too fast.
Either way, Qwen, when exactly are you going to release the official versions of these two models?
Sigh.
This morning, I found that Arena had already published the latest results for Qwen3.7 Preview.
Qwen3.7-Max-Preview ranks #13 overall in text.
Ahead of it are Claude Opus4.6 and the 4.7 series, Gemini-3.1/3 Pro series, GPT-5.4/5.4-High, and more.
As you can see, Qwen3.7-Max-Preview is also the only Chinese model in Arena’s text leaderboard top 15 worldwide.
At the same time, it also shows strong performance in subdomains, placing in the global top 10 in the following tracks:
- Math: #7
- Expert Prompts: #9
- Software/IT: #9
- Coding: #10
Another set of results released at the same time — Qwen3.7-Plus-Preview — also performed impressively in vision.
It ranks #18 in its field, lifting Alibaba’s overall ranking to #5.
It is likewise the only Chinese model listed in the vision leaderboard.
In addition, Arena also released the Expert Arena results.
In the Expert Arena, Qwen3.7-Max-Preview ranks #9 in the expert-prompt-only category.
There is another Chinese model on this leaderboard: Xiaomi’s Mimo v2.5 Pro, which ranks #7.
What more can we say?
Only this: hurry up and release the official version! Hurry up and open source it! I want to use it!
Since entering the Qwen3 era, model iteration has clearly accelerated
Some网友 said that Qwen’s new-model launches are something else: people haven’t even finished discussing one model before the trailer for the next one is already auto-playing.
This is not just a feeling. We plotted Qwen’s release and iteration timeline and found that this is indeed the case.
(Note: the table mainly includes major milestone versions; there are actually many more sub-sizes and specialized models, such as Coder, VL, Omni, and so on.)
Looking at this table, in the early period (2023–2024), Qwen’s release pace was fairly in line with the industry norm, with a major updated version roughly every 4–6 months.
The turning point came with the Qwen3 series.
Qwen3, released in 2025, includes both dense and MoE models, with parameter sizes ranging from 0.6B to 235B, supporting both complex reasoning and fast-response modes.
Among them, Thinking mode (for complex reasoning) is mainly intended for complex reasoning, long-horizon decision-making, and agent tasks; Non-Thinking mode (for fast response) focuses on low latency and quick replies.
The Qwen3.7 Preview released so far is the latest achievement in this family.
We noticed that from Qwen3 onward, the iteration pace has clearly accelerated, with major version intervals shrinking to every 2–3 months.
Especially in 2026, Qwen has had new moves almost every month. From 3.5 to 3.6 and then to 3.7, the version numbers have been updated very quickly.
What’s more, the usual pattern is for the Preview version to go live first — the community tests it / developers start using it — and only then does the official version follow.
To some extent, this reflects that Alibaba’s Qwen team has entered a phase of rapid experimentation and high-frequency delivery.
I still remember the tweet posted by Qwen’s former head, Lin Junyang, when he left:
Qwen brothers, just keep doing what was originally planned — no problem.
When he left, Qwen was in the 3.5 era, and, like DeepSeek, was a benchmark model in China’s open-source community.
Whether it’s a continuation of the old style or the guidance of the new lead, the subsequent Qwen3.5 and 3.7 versions have indeed kept moving faster — and getting better — step by step.
In the past, we waited years or even a year for a new phone upgrade; now, within just months or even weeks, we watch a model go from immature to powerful.
Some cheer, some feel anxious, and some debug through the night, all just to keep up with the relentlessly advancing curve of intelligence.
Technology never waits for anyone.
It is like a surging river, and we stand on the bank — both witnesses and participants.
References: