Qwen3.7-Plus: The Revolution of Precise Clicking for $0.40

Qwen3.7-Plus, an innovation from Alibaba, has demonstrated its ability to accurately identify the exact pixel to click on a messy screenshot of the AWS console. By simply asking where to click to launch an instance, the model provided the precise coordinates: (x=1147, y=283), located at the center of the "Launch Instance" button. This feat is offered at a cost of $0.40 per million input tokens, which is one-sixth the rate of the Qwen3.7-Max model, which is text-only.

The Qwen3.7-Plus model stands out with a score of 79.0 on ScreenSpot Pro, a key benchmark for assessing the effectiveness of "computer usage" agents. This success is based on the concept of GUI grounding, where the model, given a screenshot and an instruction, must determine the exact pixel coordinates corresponding to the targeted user interface element.

Practical Applications of Qwen3.7-Plus

The article highlights four concrete uses of the model:

1. From screenshot to JSON coordinates: The model converts screenshots into actionable data.
2. Coordinates to real clicks: With a confidence gate, the coordinates are transformed into click actions.
3. Observation-action loop with Playwright: To automate browser tasks, the model executes observation and action loops.
4. Screenshot to code: This feature allows for the recreation of user interface components from images.

Implementation and Comparison

To implement Qwen3.7-Plus, Alibaba offers the use of its Cloud Model Studio via the OpenAI-compatible SDK, thus facilitating the integration of the model into various systems. In comparison, Qwen3.7-Plus outperforms Qwen3.7-Max and other benchmarks in terms of accuracy and cost, making it an attractive option for developers looking to prototype advanced screen grounding solutions.

Limitations and Opportunities

Although Qwen3.7-Plus is a proprietary solution accessible only via API, with no possibility for self-hosting, it represents a cost-effective option for prototyping advanced screen grounding. This approach allows for testing concepts before moving on to more sophisticated solutions, whether managed or self-hosted.