Friday brought Qwen 3.8 27B, the latest Apache 2 licensed model from Alibaba’s Qwen research lab. At 27 billion parameters, it’s small enough to run on a decent laptop but claims benchmark scores that put it ahead of much larger models, including its own predecessor Qwen 3.6 27B and even the closed-weight Qwen 3.7-Plus.

The benchmarks look impressive on paper. The real question is whether it’s actually good to use.

Simon Willison spent the weekend with it and found a model that’s genuinely capable but has one very specific problem: it defaults to overthinking everything. Ask it a straightforward question and you’re likely to get a multi-paragraph meditation on the nature of the question itself before it gets around to answering.

This isn’t a deal-breaker. It’s controllable. You can tune the behavior with system prompts or temperature adjustments. But it’s worth knowing upfront that the default experience involves a lot more verbosity than you probably want.

What makes 27B the right size

There’s something interesting happening at this model size. 27B is big enough to be genuinely useful but small enough to run locally without exotic hardware. That’s increasingly important as developers realize they don’t always need frontier-class models for every task.

Qwen 3.6 27B was already strong. If the benchmarks hold up in practice, Qwen 3.8 27B represents a meaningful step forward in that sweet spot between capability and practicality.

The vision capabilities are notable too. Multimodal support in a model this size makes it more versatile than text-only alternatives, especially for local deployment scenarios where you want one model that can handle different input types.

The overthinking problem

The verbose default behavior isn’t random. It appears to be an artifact of the training approach, possibly optimized for benchmark tasks that reward thorough reasoning. That’s fine for some use cases. It’s annoying for others.

The good news is that it responds to prompt engineering. Tell it to be concise and it can be. Adjust the temperature and you’ll see different behavior. But you shouldn’t have to fight your model to get straightforward answers to straightforward questions.

This is a solvable problem, either through fine-tuning or by wrapping the model in a layer that adjusts its behavior based on the query type. But out of the box, expect to do some tuning.

What this means for local AI

The real story here isn’t just another model release. It’s that the gap between cloud-scale models and locally-runnable models keeps shrinking. A 27B model that can hold its own against much larger alternatives changes the calculation for when you actually need to hit an API.

Apache 2 licensing matters too. You can modify it, deploy it commercially, build products on top of it. That’s different from “open weights” models that let you run them but not much else.

If you’re building something that needs an LLM but doesn’t need GPT-4 class reasoning, this is worth testing. Just be ready to spend some time tuning the verbosity down to something manageable.

The model is available now through the usual channels. Expect to see it in Ollama, LM Studio, and other local inference tools shortly if it’s not there already.