Article · 2026-08-16 · 6 moments

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

Friday's big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor Qwen 3.6 27B was impressive. Qwen's self-reported benchmarks for this model are eye-opening. They show a boost from both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which was one of Qwen's strongest models of any size as recently ✦ AI generated

Read original ↗

HOST Simon Willison

Context

A 17GB open-weight model running on a consumer laptop is now competitive with the best proprietary models from a year ago, demonstrating rapid progress in local LLM capability.

The author considers it remarkable that a 17GB model can run locally and match what the most expensive proprietary models could do just a year ago. The only remaining limitation is inference speed.

transcript

Simon Willison: The fact that a 17GB file can do all of this stuff on my home machines is a miracle. Once again, I'm delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models - today it can run on a capable laptop.

gives example · 1

Example◆

Even the simplest prompt—drawing a circle as SVG—triggers minutes of elaborate internal deliberation and produces an over-engineered animated result instead of the straightforward output requested.

When asked to draw a simple SVG circle, the model spends minutes thinking about palette choices, animation, and concentric guide rings before producing an elaborate animated circle that was never requested.

transcript

Simon Willison: Qwen's reasoning trace started like this: The user is asking for an SVG drawing of a circle. Simple request — but I want it to be a carefully crafted piece. Let me make something that goes beyond just <circle>: a single self-contained SVG file with character — maybe a geometric "circle study," with subtle animation, layered rings, and a distinctive palette. [...] Several minutes later it produced this absolutely beautiful animated circle, which was entirely not what I had asked for!

Mechanism

Multi-Token Prediction provides a significant inference speed boost for Qwen 3.8 27B, with one test showing a 72% improvement over the default GGUF configuration.

The model runs slowly by default (15-30 tokens/second) but supports Multi-Token Prediction, a built-in optimization that can boost speed dramatically—tested at 72% faster on a DGX Spark.

transcript

Simon Willison: One of the most promising optimizations is baked into the model itself. Qwen supports Multi-Token Prediction, an architecture trick where a cheaper mechanism guesses several tokens ahead and the main model can then quickly verify if the guesses were correct. This can have quite a dramatic effect on inference performance. I had GPT-5.6 in Codex run a comparative benchmark on the Spark and the --spec-type draft-mtp server outperformed the LM Studio default GGUF by around 72%.

Fact

Qwen 3.8 27B has sufficient capability in long context, code generation, and tool-calling to successfully drive a coding agent loop on consumer hardware.

Local experimentation with the Pi coding agent showed promising results. The model analyzed a codebase and built a working utility tool to convert session transcripts to markdown.

transcript

Simon Willison: One of the biggest questions around local models is whether or not they have enough horsepower to successfully run a coding agent loop. Coding agents require long context, strong code generation support and reliable tool-calling. On paper Qwen 3.8 27B has all three of these, so is it up to the task? My initial experiments with Pi have been very promising.

supports · 1

Fact

Qwen 3.8 27B is highly accurate at generating bounding box coordinates for objects in photographs, producing near-perfect results on its first attempt.

The model produces accurate 0-1000 scale bounding boxes for objects in photos. A test with pelicans showed near-perfect alignment when rendered on the original image.

transcript

Simon Willison: I've been asking for 0-1000 scale to produce good results in the past. I tried this: Return JSON bounding boxes for the pelicans in this photo, 0-1000 scale for each dimension. This is such a good match. Here are those boxes rendered on top of the photo.

Claim◆

Qwen 3.8 27B's default xhigh reasoning setting causes the model to dramatically over-think even trivial prompts, wasting tokens and time.

The model defaults to xhigh reasoning, which causes absurd over-thinking on simple tasks. The author strongly recommends using low or no reasoning to start.

transcript

Simon Willison: Qwen's documentation describes the model as defaulting to xhigh for the reasoning effort, and the LM Studio GGUF I've been trying preserves that default: xhigh (default): for complex tasks demanding thorough analysis [...] This is a hilarious default. It's absolutely not a good way to run the model, especially on consumer hardware. [...] My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.

gives example · 1

Highlight slides

Simple prompts trigger elaborate deliberation✦ from: Even the simplest prompt—drawing a circle as SVG—triggers minutes of elaborate internal deliberation and produces an over-engineered animated result instead of the straightforward output requested.Qwen 3.8 27B: Default xhigh Reasoning Problem✦ from: Qwen 3.8 27B's default xhigh reasoning setting causes the model to dramatically over-think even trivial prompts, wasting tokens and time.Reasoning trace reveals over-ambition✦ from: Even the simplest prompt—drawing a circle as SVG—triggers minutes of elaborate internal deliberation and produces an over-engineered animated result instead of the straightforward output requested.Simon Willison's Recommendation✦ from: Qwen 3.8 27B's default xhigh reasoning setting causes the model to dramatically over-think even trivial prompts, wasting tokens and time.

Related episodes

Video · The Cognitive Revolution

Fable's Back, AI Engineer Recap, & SambaNova

→

Video · Lenny's Podcast

The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)

→

Article · Big Technology

Everything You Need to Know About Kimi K3, the Latest Model From China to Shake Up AI

→