Qwen3.7-Max Enters Global Top 4 in Code Arena, Challenges Claude

Today, the latest Code Arena leaderboard has been released!

Qwen3.7-Max scored 1,541 points, breaking into the global top four and surpassing leading models such as GPT-5.5 and Gemini 3.5 Flash.

Only Claude Opus 4.7 and Opus 4.6 rank above it.

In other words, on the global programming model stage, Alibaba is the only Chinese company to make it to the table, ranking second behind Anthropic.

Actually, before the Code Arena rankings were released, Qwen3.7-Max had already gained a reputation among overseas developers.

Atomic Chat conducted a head-to-head comparison, pitting Opus 4.7, GPT-5.5, and Qwen3.7-Max against each other in a challenge to write a Tetris AI capable of self-training.

As a result, Qwen3.7-Max not only surpassed Opus 4.7 and GPT-5.5 with just $1.32 in token costs, but also improved performance by 56%.

Another overseas developer chose to use Qwen3.7-Max to build a 3D model of the universe, with results that can only be described as breathtaking.

In the generation task for the "3D Pixel-Style Mini Pagoda Model," Qwen3.7-Max outperforms in both speed and quality.

Developer Paul Couvert further praised that, after integrating Hermes Agent and OpenCode, Qwen3.7-Max can essentially replace GPT-5.5 and Opus 4.7.

Programming, too powerful

No matter how high the benchmark score, nothing beats real-world testing.

We put Qwen3.7-Max through a rigorous "racing game" challenge.

After inputting a detailed prompt, Qwen3.7-Max immediately generated a playable HTML file.

The first version had a small bug: the A/D steering keys were reversed.

But after a second round of simple dialogue fine-tuning, a fully immersive 3D racing game was up and running.

At the moment it opened, I have to admit, I was a bit surprised.

Four cars compete on the same track, racing around a circular circuit three laps long, with over 100 gold coins scattered across the track; hitting obstacles will slow you down and cause loss of control.

Post-race results panel with rankings, time, coin count, and fastest lap—everything you need.

But what's truly surprising are two details that only Qwen3.7-Max accomplishes.

One of them has a start screen. After testing all four models horizontally, only this one provided a proper start page for the game—you must click “Start” to begin the match. The other three launch directly into the game with no title screen at all.

The other one is the sound effects. The prompt ended with a request to include engine revving and coin-collecting sounds. Among the four models, only this one incorporated that bonus—both the engine noise and the coin chimes were added.

Take a look at the performance of the other participants.

The Gemini 3.5 Flash image appears noticeably flatter, lacking that vivid, three-dimensional depth.

The UI layout also has issues—the dashboard information is scattered across the four corners of the screen, resulting in a fragmented visual focus.

In comparison, Qwen3.7-Max centers the key metrics on the screen, better aligning with the natural focus of the player's gaze.

The performance of Claude Opus 4.6 is somewhat hard to put into words.

Not only are there barely any coins on the track, but the three AI cars drive in perfect sync, with no randomness—they look like copies of each other.

Finally, GPT-5.5.

You can see that the visual quality is significantly better than the previous two, and the operation is also smoother.

But for some reason, the coins were made into yellow "donuts"...

The design is minor; the key point is that Gemini, Claude, and ChatGPT all went through multiple rounds of bug fixes to get all features working properly.

Only Qwen3.7-Max is playable right from the first generation.

Performance is nearly identical, real-world tests confirm it, and the price is only a fraction. The rest is up to developers to decide with their actions.

Qwen3.7-Max achieves such impressive performance on the most competitive programming battlegrounds because its product positioning holds the answer.

A few days ago, when Alibaba released Qwen3.7-Max, it assigned it a very special label: Agent Foundation Model.

It was designed from the start to autonomously execute tasks over extended periods.

Internal test data shows that during a single autonomous programming task, Qwen3.7-Max ran continuously for 35 hours and executed 1,158 tool calls.

The final generated code achieves an impressive 10x geometric mean speedup compared to the Triton reference implementation.

Even more impressive is its "endurance" capability—

After 30 hours of simulation, the model remained sharp, continuously uncovering new areas for optimization.

Zero context degradation, zero instruction drift, zero infinite loops!

To be honest, the challenge isn't in making 1,000 tool calls itself. After the MCP protocol is widely adopted, making 1,000 tool calls won't be unusual.

The challenge lies in 35 hours of continuous reasoning.

Most models crash when running long tasks: either the context becomes increasingly disorganized, causing them to completely forget the goals set at the beginning; or they enter an infinite loop, repeatedly attempting the same failed approach.

Qwen3.7-Max has successfully implemented the principle of "consistently doing the right things."

Core Technology Revealed

This leap in programming capability with Qwen3.7-Max is likely related to upgrades in two training methods.

The first is environment expansion.

During programming training, Qwen3.7-Max breaks each task into three independent dimensions: the task itself, the execution framework, and the validation method, which can be freely combined.

The same question is sometimes solved within the Claude Code framework, sometimes within OpenClaw, and sometimes with a different verification method.

It’s like an intern rotated through every project team—what it’s forced to learn are general problem-solving strategies, not "how to game a specific framework."

This explains an counterintuitive phenomenon: Qwen3.7-Max performs consistently across frameworks such as Claude Code, OpenClaw, and Qwen Code, without showing a pattern of being strong in its own framework but underperforming in others.

The second upgrade is long-range autonomous execution.

During training, the team introduced the "Dynamic Accumulative Survival Game" framework.

That is, having the model make over a thousand consecutive decisions in a continuously changing simulated environment, independently forming hypotheses and adjusting strategies based on feedback, without suffering from "context degradation" due to prolonged execution.

Here is an intuitive data point: In the YC-Bench simulation of a startup operating for a full year, Qwen3.7-Max achieved $2.08 million in revenue, doubling the previous generation’s $1.05 million.

More importantly, it demonstrates strategic evolution: during mid-term crises, it can autonomously adjust its direction, identify and block malicious clients, and ultimately converge into a stable execution cycle.

This is the underlying support for the 35-hour kernel optimization case, and it's also why Qwen3.7-Max achieves acceleration in 96% of scenarios on Kernel Bench L3.

But programming is only the first battlefield. The foundation of this long-range reasoning toolset points to a greater ambition—a general-purpose Agent base.

Since its launch, Code Arena has always tested real skills—multi-step reasoning, tool orchestration, and complete project delivery—all true Agent-level challenges.

Today, Qwen3.7-Max scored 1541 points, securing fourth place between Opus 4.6 Thinking and Opus 4.6.

On a track where Claude has dominated for over a year, it has provided its own answer: Chinese models are not just followers—they can also be definers.

The global programming model competition is no longer just a solo act by Silicon Valley.

Reference materials:

https://arena.ai/leaderboard/code/webdev

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.