Qwen3.8 27B impresses but defaults to "overthinking"

Qwen3.8 27B is a 27-billion-parameter vision-capable language model released under the Apache 2.0 license by Alibaba's Qwen research lab. In the 27B class — a size that runs comfortably on a single gaming laptop — its predecessor, Qwen3.6 27B, had earned strong reviews, so the follow-up drew significant attention. According to Alibaba's own benchmark disclosures, the new model reportedly outscored not just its predecessor but also Qwen3.7-Plus, the closed-source model that was once the top of the Qwen lineup. However, these figures came from Alibaba itself, and independent verification has not yet been confirmed.

The problem: a default that "thinks too much"

Willison tested the model using LM Studio on two machines: a MacBook Pro with an M5 Max chip and 128GB of memory, and an NVIDIA DGX Spark. The actual build he ran was a Q4_K_M quantized version taking up 17GB of disk space, and on the DGX Spark he also tried running llama-server directly.

The root cause turned out to be a configuration issue. Qwen's official documentation lists "xhigh" as the default reasoning_effort setting for this model, and the GGUF build distributed by LM Studio inherited that same default. Combined with LM Studio's default context length limit of 8,192 tokens, the model ended up spending its entire remaining token budget on reasoning even for something as simple as "draw a circle." Willison was only able to avoid this by raising the context length to the model's maximum supported value of 262,144 tokens.

The circle-drawing test illustrates this most clearly. With the xhigh default active, the model's reasoning process began with "The user is asking for an SVG drawing of a circle" and then spent several minutes adding concentric rings, compass-like tick marks, and subtle animation on its own initiative. The result was attractive, but it wasn't the single circle that had actually been requested.

Reasoning on vs. off comparison

Category	Reasoning on (xhigh default)	Reasoning off
Time taken	21 minutes	137 seconds (about 2 min 17 sec)
Reasoning tokens	22,276	N/A
Output tokens	3,223	3,715

The version with reasoning turned off didn't produce exactly the requested image in one shot, but it delivered an answer within two minutes. By contrast, in a bounding-box test — a vision task that requires pinpointing the coordinates of an object in a photo — the version with reasoning enabled placed the box correctly on a pelican photo, while the version with reasoning disabled placed it incorrectly. In other words, the reasoning isn't purely wasted effort.

Also tested as a coding agent

Willison connected Qwen3.8 27B to Pi, an agent tool whose short system prompt makes it well-suited to smaller models. He ran the model via LM Studio on the DGX Spark, shared it using tailscale serve, registered it in the ~/.pi/agent/models.json config file, and had it work on tasks in an actual development folder. The model moved across multiple files, alternating between reasoning and tool calls, and eventually produced a usable answer. It then went on to write and successfully run a Python script that converts conversation log files (JSONL) into markdown. In other words, a 17GB local model handled most of the functions needed for real work — writing code, calling tools, and annotating images.

Editor's view

What makes this review interesting isn't a performance flex — it's a precise diagnosis of a "default-setting trap." Alibaba's documentation officially states that the reasoning_effort option can be adjusted. But the actual GGUF build that got distributed shipped with the heaviest possible default, xhigh, and most ordinary users downloading it have no idea, ending up burning minutes just to draw a circle. It's a reminder that what holds open-source models back is often not the model's underlying capability but the default settings chosen at deployment time.

Experience running 27B-class local models in real workflows tends to point to the same conclusion every time: leaving a reasoning-enabled model at its default setting slows down perceived performance on simple tasks, while for complex vision or coding work, that same reasoning actually boosts accuracy. In practice, that means developers need to get in the habit of dialing reasoning intensity up or down — or off entirely — depending on the task at hand. The fact that the Qwen3.8 line has been sitting near the top of Hugging Face's trending charts and local-inference rankings since its release in early August isn't unrelated to this flexibility.

From a practical standpoint, this model is worth trying for teams looking to run code generation, image annotation, or simple agent tasks on a personal laptop or small workstation. But before deploying it, teams should check the context length and reasoning_effort defaults and adjust them as needed. Skipping that step before putting it into a live service risks noticeable response delays.

In the coming weeks, deployment tools like LM Studio and Ollama are likely to respond to feedback like this by lowering their default reasoning_effort settings or making the setting more explicit to users. At the same time, independent benchmarking groups are expected to publish results verifying Alibaba's self-reported scores for Qwen3.8 27B before long.