Qwen3.7 dual model sneaks onto Arena AI, official release imminent

On May 19, Alibaba Cloud quietly released two new models on Qwen Chat and Arena AI (formerly LMArena): Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview. No launch event, no blog post—barely even a tweet, just a casual “Qwen 3.7 Preview.” But everyone knew what that meant—tomorrow (May 20) is Alibaba Cloud Summit, and the official release of Qwen3.7 is locked and loaded.

This “drop results first, launch later” playbook is now second nature for the Qwen team. From Qwen3-Max to Qwen3.5, then to Qwen3.6, every flagship generation has followed the same pattern—get scores on Arena AI first, then do the official release. The reason is simple: posting strong scores on a third-party blind-test leaderboard is far more convincing than writing your own tech blog.

Leaderboard Results: Solid, But Not a Runaway Lead

Let’s start with the hard data. On Arena AI’s text arena, Qwen3.7-Max-Preview ranks 13th overall, putting Tongyi Qianwen’s lab at 6th place among all participants. Its performance across sub-leaderboards is worth a closer look:

• Math: 7th
• Expert Applications: 9th
• Software & IT: 9th
• Programming: 10th
• Expert Arena (expert prompts only): 9th

On the vision side, Qwen3.7-Plus-Preview ranks 16th overall, lifting Tongyi Qianwen’s lab to 5th place on the visual leaderboard.

So what does this mean? Frankly, it’s not earth-shattering—but it’s far from weak. GPT, Claude, and Gemini currently dominate the Arena AI top ranks, and the top 10 is fiercely contested. For Qwen3.7-Max-Preview to break into the top 13 and stay consistently within the top 10 for math, programming, and expert tasks signals that the new flagship definitely isn’t a minor tune-up.

What’s most noteworthy is the 9th place in the “Expert Arena.” This sub-board involves prompts submitted by advanced users—developers and researchers posing real-world, complex problems. Scoring in the top 10 here means the model truly has strength in long reasoning chains and complex task decomposition.

Preview Versions Show Restraint: Thinking Mode Only, All Tools Disabled

Both Preview versions share the same feature limitations:

• Supports Thinking Mode only
• Search tool disabled
• Code interpreter disabled

This tradeoff is typical. During the Preview phase, isolating the core model and stripping away external tools keeps leaderboard evaluations clean—no chance of “the model performs poorly but search saves it.” Under Arena AI’s blind-test mechanism, tool calls can introduce noise, so closing them off is the more honest choice.

As for limiting to Thinking Mode—based on Qwen3.5’s design philosophy, the official release will likely continue using the “Thinking / Non-thinking dual-mode fusion” architecture. Historically, the Plus series balances performance, speed, and cost, while the Max series aims for top scores and heavy-duty tasks. That division of roles will probably continue here.

From 3.6 to 3.7: Iteration Speed Is Unusually Fast

One thing worth noting is the pace of iteration. Alibaba Cloud’s update frequency has visibly accelerated this year:

• Qwen3.5 series emphasized long context (up to 1 million tokens) and multimodal fusion
• Qwen3.6-Max-Preview focused on “stronger world knowledge, instruction following, Agentic Coding”
• Qwen3.6-Plus improved Vibe Coding and multimodal recognition
• Now, Qwen3.7 dual Previews have arrived

Major version intervals have shrunk from months to mere weeks. Behind this are two factors: more mature training pipelines and pressure from competition—Anthropic, Google, and OpenAI are all iterating fast; domestically, DeepSeek V4 is already out in Pro and Flash variants—Qwen simply can’t afford to move slowly.

From available information, Qwen3.7 likely continues to focus on three key areas:

1. Agentic Coding: This has been Qwen’s main emphasis since 3.6; its 10th place in the Programming sub-board shows clear ongoing investment.
2. Math and Reasoning: The 7th place in math is the standout result.
3. Multimodal: The Plus-Preview’s strong vision score shows its VL capability hasn’t fallen behind.

Why This Release Matters for Developers

To be blunt, Arena AI rankings are just reference points. What truly determines whether a model can be used in production are several practical factors:

First, pricing. Qwen3.6-Max-Preview’s tiered pricing is 9–15 RMB per million input tokens and 54–90 RMB per million output tokens. If the Qwen3.7 official release maintains that range, it’ll stay an excellent value domestically. Compared with flagship prices from GPT and Claude, Qwen-Max remains the “good enough and affordable” choice.

Second, context length. Qwen3.5-Plus already supports a 1 million-token context, while the Max series stays around 260k. If Qwen3.7 increases Max’s context to the million range too, long-document and code repository analysis will become much smoother.

Third, Agent capability. In modern Agent applications, the hardest part isn’t reasoning—it’s unstable tool calls and long task chains going off track. Qwen3.6 made big strides in Agentic Coding, and if 3.7 continues to improve here, it’ll be a major benefit for domestic teams developing coding agents and automated workflows.

May 20 Summit Preview

As usual, tomorrow’s Alibaba Cloud Summit should see the official release of several things:

• Qwen3.7-Max official version (enabling Thinking/Non-thinking modes, search, and full code interpreter functionality)
• Qwen3.7-Plus official version (complete multimodal capabilities)
• Most likely Qwen3.7-Flash, targeting low-cost, high-speed scenarios
• Associated API pricing, context specs, and tiered billing details
• Potential upgrades to Qwen Chat’s product interface

The open-source version is especially worth anticipating. The Qwen series has been one of the most generous domestic open models—both Qwen3 and Qwen3.5 have open-source releases on Hugging Face. If Qwen3.7 continues this tradition, the community ecosystem will heat up again.

By the way, OpenAI Hub is already preparing to integrate the Qwen3.7 series. Once the official versions launch, developers will be able to use a single key to switch freely between GPT, Claude, Gemini, DeepSeek, and Qwen—making multi-model comparison and routing much easier, especially for teams already running A/B tests across different models.

Final Thoughts

China’s large-model race heading into 2026 is no longer about “can we build it,” but “can we keep up with the iteration pace.” Qwen’s strategy over the past two years has been clear: release leaderboard results first, products second; keep prices down; follow up with open-source. This playbook is still in action for 3.7.

Rank 13 isn’t the endpoint—the official version launching tomorrow is. If Qwen3.7’s full-featured release can climb higher once tools are re-enabled, the landscape of domestic flagship AI models may be reshuffled yet again.

References

• ITHome: Alibaba Cloud’s Qwen3.7-Max-Preview Debuts on Arena AI — initial coverage with detailed Arena AI ranking data
• Zhihu Column: A New King of Chinese AI! Qwen3.6-Max-Preview Appears — in-depth analysis of the previous generation Qwen3.6-Max-Preview, useful for comparison