On 16 August 2026 Alibaba released Qwen 3.8 27B — an open, multimodal language model under Apache 2. Simon Willison rated it very good but warns: by default it sets the highest reasoning level and burns compute even on trivial prompts.
Key takeaways
- Qwen 3.8 27B: 27 billion parameters, multimodal, Apache 2 license.
- Default reasoning level is "xhigh" — a poor choice for consumer hardware, per Willison.
- The reasoning_effort parameter has three levels: low, medium, xhigh.
- Context window up to 262,144 tokens.
- Multi-Token Prediction optimization gives about a 72% speedup.
Overthinking by default
The biggest surprise isn't quality, it's the default. The model starts with reasoning_effort at "xhigh," so even a simple prompt can trigger a long deliberation. Willison describes how a request to "draw an SVG of a circle" set off elaborate musings about aesthetics and ended with an animated circle instead of a plain one. He calls the default "hilarious" and openly advises against it on consumer hardware.
Ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.
Simon Willison.
| Level | When to use |
|---|---|
| low | simple tasks, minimal spend |
| medium | balanced trade-off |
| xhigh | default — maximum reasoning, high cost |
How to run it and for whom
The model is about 17 GB in Q4_K_M quantization, which fits on beefier laptops. Willison tested it on an M5 Max MacBook Pro (128 GB) and an NVIDIA DGX Spark, using LM Studio and llama-server. Speed is 15–30 tokens per second — slower than hosted APIs, but fine for local work. He showed SVG generation, bounding-box detection, HTML and Python code, and use as a coding agent.
The context window reaches 262,144 tokens, and the built-in Multi-Token Prediction optimization gives about a 72% speedup.
on DGX Spark: llama serve \ -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \ -hfd ggml-org/Qwen3.8-27B-GGUF:Q4_0 \ --spec-default \ --spec-type draft-mtp \ --reasoning-preserve --agent
Why it matters
An open 27B model that fits on a single strong laptop and handles code and images lowers the barrier to local AI.
But the "xhigh" default points to a wider issue: vendors increasingly ship maximum reasoning as the default, which drives up cost and latency without asking the user. Deliberately managing the reasoning budget becomes a real skill, not a detail — especially when the model runs locally and you pay with your own power and time.
What's next?
- Local users should explicitly set reasoning_effort to low/medium to avoid wasting compute, per Willison's advice.
- The open Apache 2 license allows commercial deployment and fine-tuning without licensing limits.
- A real test will be comparing Qwen 3.8 27B with other open models of its class on coding tasks.
Sources
- Simon Willison's Weblog — Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
- Hugging Face — Qwen (organization, model cards)
