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

Qwen 3.8 27B is a highly capable, open-weights, vision-enabled LLM (17GB file) that runs well on consumer hardware, demonstrating remarkable progress in local models.
Its default reasoning_effort: xhigh setting causes “spectacular over-thinking,” leading to excessively long generation times and overly complex or unintended outputs (e.g., animated circles for a simple circle request); users should immediately adjust this to “low” or “no reasoning.”
Despite the default overthinking, it excels at vision tasks (accurate bounding boxes) and can effectively drive coding agents like Pi, showcasing strong code generation and tool-calling.
While inference speed is a bottleneck (15-30 tokens/second), the model supports Multi-Token Prediction (MTP) which can significantly boost performance (e.g., 72% faster with llama serve).