Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Why is this worth discussing? Alibaba's Qwen team released Qwen 3.8 27B, an open-source model that performed impressively in benchmarks, even surpassing their own previous closed-source models. That alone is noteworthy. But more interesting is Simon Willison's hands-on experience, which highlights a deeper issue: the massive gap between a model's default configuration and actual user needs.

The Core Issue: Is the default "deep thinking" genius or foolish? The model defaults to the xhigh reasoning depth. This means even a simple task like "draw a pelican riding a bicycle as SVG" triggers 22,000 reasoning tokens and 21 minutes of "deep thought." The result? A beautifully detailed SVG, but at a time cost completely unacceptable for a local 17GB model.

It's like asking a genius mathematician to buy groceries, and they spend 30 minutes at checkout mentally calculating the optimal discount combinations. The result is precise, but utterly unnecessary.

What trend does this reveal?

1. A model's "capability" is decoupling from its "usability." Research teams chase higher benchmark scores by defaulting to the most resource-intensive features. For average users, especially on consumer hardware, this can be a disaster. It's no longer about "how strong is the model," but "how dumb is the default configuration."
2. Users must become model "trainers." The model offers three reasoning depth levels (xhigh, medium, low). This means using a local model well requires not just hardware, but knowledge of "which setting to use when." It's a new cognitive burden.
3. The "sweet spot" for local deployment is shifting. Running 27B parameters on a laptop is cool, but if a default config makes a task take 21 minutes, its practical value plummets. This forces users and tool developers (like LM Studio) to manage context length and reasoning parameters more intelligently.

What's in it for you?

• If you're trying this model: Immediately set the reasoning depth to medium or low manually. Use xhigh only for complex analysis tasks where you're willing to wait a long time.
• If you're a developer: This is a reminder that when integrating any new model, never blindly trust default parameters. Test latency and quality trade-offs under different configurations and provide users with simple toggles. n- The bigger picture: AI model "usability" is becoming the next battlefield. Whoever offers smarter default configs and more automated parameter tuning will win the hearts of local deployment users. After all, users want tools that work, not lab products that need constant tweaking.

Simon Willison's article uses a vivid example to show us: A model's worth lies not only in how high its ceiling is, but also in how foolish its default state is.