[DigitalToday reporter Chi-gyu Hwang] Alibaba has open-sourced its multimodal AI model, Qwen3.8-27B, which is sized to run locally on higher-spec MacBook Pros or Mac Studios.

By Alibaba’s own benchmarks, it delivers performance on par with Anthropic’s Opus 4.6 Max, drawing attention.

Simon Willison (사이먼 윌리슨), who developed the data exploration tool Datasette and co-developed the Python web framework Django, published a review after using Qwen 3.8 27B himself.

Qwen 3.8 27B is a 27B AI model, meaning it has 27 billion parameters, a unit indicating how much information a model has learned. Its size means it can run on a typical laptop. It can also recognise images and is provided as free open-source software.

Willison said the model’s performance itself is impressive. But he said the default settings are a problem.

He said Qwen 3.8 27B goes through an internal thinking process before producing an answer, and the default is set to "xhigh" deep thinking. That leads it to spend longer than necessary even on simple questions, he said.

Willison said, "When I asked it to draw a pelican riding a bicycle, the model produced a single image after thinking for 21 minutes. The result was quite detailed. The bicycle frame shape was correct, and the pelican’s legs were properly drawn on both sides. But 21 minutes was too long."

When he made the same request again with the deep-thinking function turned off, it produced an answer in about 2 minutes. The image was less complete.

Willison also tested it on finding an object’s location in a photo. When he provided a pelican photo and asked for the coordinates of where the pelican was, the model found accurate coordinates. He also told it to build a tool to verify the result, and it added an example-image function on its own, even though he did not ask for it.

He said it also carried out the role of a so-called coding agent, writing code and directly modifying source-code files, without problems. Willison said that when he provided a real project folder and asked it to explain how the login function works, it searched out and read multiple files on its own and then produced an accurate answer.

But he said the speed was disappointing. "Even running it on a high-performance laptop and dedicated equipment, the response speed is on the slow side," he said. "It feels even slower, especially when the deep-thinking function is on." He added, "Fortunately, within two days of its release, the developer community has already found a way to boost speed by more than 70 percent."

The overall assessment is roughly as follows.

Willison said, "It’s surprising that a single 17 gigabyte file can do this much," adding, "Even a year ago, this level of performance was only possible with expensive commercial AI, but now a single laptop is enough." He added, "If you just lower the default settings, this model is among the best local AI available now."