AIToday
Large Language ModelsAI Coding AssistantsOpen-Source AISimon Willison's WeblogPublished: Aug 17, 2026, 10:03 JST3 min read

Qwen 3.8 27B defaults to overthinking—change that setting first

Qwen 3.8 27B defaults to overthinking—change that setting first

Key takeaway

  • Alibaba's Qwen research lab released Qwen 3.8 27B on August 16, 2026, a 27-billion-parameter open-source language model that runs on consumer hardware and supports vision tasks, code generation, and image annotation.

  • The model ships with a default "xhigh" reasoning setting that causes excessive overthinking even on simple prompts—users should switch to "low" or disable reasoning to avoid multi-minute delays.

  • While the 17GB model delivers impressive capability (vision bounding boxes, coding agents, tool use) on offline machines, inference speeds of 15–30 tokens per second lag behind cloud APIs, and upcoming optimizations like Multi-Token Prediction could improve performance by ~72%.

3 Key Points

  1. What happened

    Alibaba released Qwen 3.8 27B on August 16, 2026, an Apache 2 licensed 27-billion-parameter vision-capable LLM that can run on consumer hardware like a MacBook Pro or NVIDIA DGX. The model defaults to "xhigh" reasoning effort, which causes it to spend excessive computation on even simple tasks—a pelican SVG took 21 minutes with 22,276 reasoning tokens, versus 2 minutes with reasoning off.

  2. Why it matters

    The 17GB model performs sophisticated work locally—coding, image annotation with bounding boxes, and running as a coding agent—without relying on cloud APIs. However, inference speed of 15–30 tokens per second on the tested hardware lags far behind hosted models like OpenAI 5.6 Sol (74 tokens/second) and 5.6 Luna (184/second), making it impractical for many real-time use cases despite strong capability.

  3. What to watch

    Multi-Token Prediction (MTP), an architecture feature built into the model, showed a ~72% speedup in early testing when invoked via llama.cpp. The author expects further optimization innovations "over the next few weeks," which could make local deployment more competitive with API-based alternatives on speed.

Ask the AI about this article →

Context & Analysis

Qwen 3.8 27B arrives at a inflection point for local open-source models: capability has reached a level where a 17GB file running on consumer hardware can rival proprietary cloud APIs in versatility, yet infrastructure constraints—specifically memory bandwidth on conventional CPUs and GPUs—still impose a speed penalty that limits practical adoption as a daily-driver replacement. The model's self-reported benchmarks show improvement over its predecessor Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus (which ranked among Qwen's strongest models in May 2026), but the article does not cite independent benchmark validation yet.

The xhigh reasoning default reveals a design tension: the model's reasoning capability is genuine and measurable—it produced a nearly-correct bounding-box tool on the first attempt where reasoning-off mode failed—but the default setting is counterproductive for consumer hardware. Users must actively dial down reasoning effort to achieve practical latency. The author's examples (21 minutes for a pelican SVG with reasoning, 2 minutes without) underscore that the model's capability does not automatically translate to usability at default settings.

The emerging optimization pathway via Multi-Token Prediction (MTP)—an architecture feature already embedded in the model, now being exposed by llama.cpp—suggests that speed constraints are not fundamental to the model design but rather artifacts of current serving infrastructure. A ~72% speedup from MTP alone hints that further tuning and community innovation could significantly close the gap to cloud API speeds over the coming weeks, potentially making local deployment more viable without sacrificing capability.

FAQ

What is the default reasoning setting and why is it a problem?
The model defaults to "xhigh" reasoning effort, which causes it to spend excessive computation analyzing even mundane tasks. A simple "draw an SVG of a circle" prompt led the model to spend several minutes over-thinking the design, producing an elaborate animated circle instead of the requested simple shape. The author strongly recommends switching to "low" or disabling reasoning entirely on first use.
How fast is inference on local hardware?
The author measured 15–30 tokens per second on an M5 Max MacBook Pro and NVIDIA DGX Spark, compared to 74 tokens/second for OpenAI 5.6 Sol and 184/second for 5.6 Luna. Multi-Token Prediction optimization showed a ~72% speedup in early benchmarks, suggesting future improvements could narrow the gap.
Can it run coding agents and tool use locally?
Yes. The author successfully configured the model with Pi (a coding agent framework) to answer questions about code, generate Python scripts, and execute them locally. It also performed vision tasks like drawing bounding boxes around objects in photographs, demonstrating reliable tool-calling and long-context support.
Simon Willison's WeblogRead Original Article

Get the latest Large Language Models news every morning

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Free · takes 30 seconds · unsubscribe anytime

Ask AI

Ask AI anything about this article. Q&As are published on this page for other readers too.

Related Articles

Next articleAmodei: AI faces trust crisis as people fear companies 'cooking up' new ways to harm them

The AI news that matters, in one minute each morning.

Sign up free